Cognitive PC Inc. · Seed Round · February 2027

The Cognitive PC.

A 15W in-memory NPU, 128GB of unified memory,
and a Linux operating system that remembers.

15W + 128GB

Dual-subsystem AI hardware

An always-on Axelera Metis D-IMC NPU (214 TOPS INT8, 3.5–9W typical) beside a 128GB unified-memory APU. Perception stays off the LLM memory bus.

0.58 ms

System-level retrieval, in Rust

hips — 34,922 lines of open-source Rust. p50 lexical BM25 across 7.1M Wikipedia articles on an Apple M3 (p95 11 ms); 2.6–14 ms hybrid code search.

$2.5M

Seed on a YC Post-Money SAFE

First outside round ($0 raised to date). Two full-time co-founders, ~100k lines of Rust across founder projects (35k open-source, assigned to the company), and a 5-node test bed. $12.5M cap.

Ilias Chrysovergis, CEO · Iason Malkotsis, CTO 1 / 12

The problem

Your PC still thinks like it’s 1984.

Cloud AI has been bolted onto CPUs and GPUs that were never designed for 24/7 perception, persistent memory, or safe local agency.

360W

The silicon wall

Discrete GPUs hit a hard 16–24GB VRAM cliff. Unified-memory PCs suffer severe bandwidth contention when background indexing competes with 70B LLM decoding.

Amnesia

The operating-system wall

Windows, macOS and Linux manage files and pixels, not semantic context. Every app spins up its own Python/Electron RAG daemon, burning gigabytes of RAM.

$250/mo

The cloud trap

Anthropic reports $150–250 per developer per month on Claude Code on average; heavy agentic users exceed $1,000. All of it while streaming proprietary code, clinical records and legal files to someone else’s servers.

Cognitive PC Inc. 2 / 12

The product

Hardware and OS, co-designed for intelligence that never sleeps.

A desktop and workstation computer where dual-subsystem AI hardware and a Linux-based OS with a Rust-native cognitive layer are built together for 24/7 autonomous intelligence and low-latency spatial computing.

Hardware

Cognitive PC Mark I

$3,500

AMD Ryzen AI Max+ 395 with 128GB LPDDR5X unified memory, paired with a dedicated 15W Axelera Metis M.2 D-IMC NPU (214 TOPS INT8, 32 MiB SRAM). Continuous ingestion runs beside, not on, the bus used for 35B–70B LLM reasoning.

Operating system

Cognitive OS

$49/mo Pro · $199/mo Ent.

A Linux-based OS and cross-platform daemon with system-level semantic memory (hips), scoped agent memory, Landlock sandboxing, and a perception path built for XR streaming at Vision Pro–class latency (12 ms photon-to-photon).

Silicon roadmap

Cognitive NPU

$599Upgrade Kit

We ship COTS workstations and M.2/PCIe NPU Upgrade Kits now, while synthesizing custom stochastic neural queueing and KV-cache compression blocks on FPGA toward a Series A silicon shuttle.

Cognitive PC Inc. 3 / 12

The hardware · Local AI · Real-world context · No compromises

Palm-sized. Datacenter-class intelligence.

AMD Strix Halo (Ryzen AI Max+ 395)

128 GB LPDDR5X Unified Memory (≈121 GiB usable)

Dedicated to LLM reasoning, 3D spatial compositing and XR workloads.

Axelera AI Metis (D-IMC)

15W AI co-processor · 214 TOPS INT8 · 32 MiB SRAM

Continuous vision, 6DoF tracking, OCR and perception — isolated from host memory.

Full I/O for real work

2 × USB4 (40 Gbps) · HDMI 2.1 / DisplayPort 2.1 · 2.5 GbE · M.2 (NVMe) · Wi-Fi 7 / Bluetooth 5.4 · 40-pin expansion (GPIO, sensors, XR).

Dual-subsystem architecture

Physically isolated compute paths for reasoning (LLM) and real-time perception (vision/XR). No memory-bus contention.

Designed for the real world

Quiet (~15–25W typical). Runs a full Linux Cognitive OS. Developer-friendly. Enterprise-ready.

Cognitive PC Inc. 4 / 12

Inside the machine

Two stages. One bus left alone.

On a Mac Studio or a standalone Strix Halo PC, background indexing steals memory bandwidth from LLM decoding. A dedicated perception stage with its own on-chip SRAM keeps that traffic off the main bus (<2% PCIe input DMA).

Stage 1 · always on

15W D-IMC NPU

Axelera Metis · 214 TOPS INT8 · 32 MiB SRAM · 3.5–9W typical. OCR, 6DoF tracking, intent routing — on-chip.

wakes only when
uncertainty > τ

Stage 2 · on demand

128GB unified memory

Ryzen AI Max+ 395 · 256 GB/s LPDDR5X. Bus reserved for 35B–70B LLM decoding and 3D compositing.

Hardware

Confidence-gated cascade. Preliminary results on a Metis → RTX 5080 cascade (manuscript in preparation, UTH COMNET Lab) show multi-× lower active energy than a standalone GPU.
FPGA → custom silicon. Hardware confidence gating and stochastic neural queueing token routing, synthesized on Xilinx UltraScale+/Versal.

Software

hips search daemon — 34,922 LOC Rust. Memory-mapped inverted index, Block-Max WAND, FP16 vector store. 0.58 ms p50 lexical BM25 on 7.1M articles; 2.6–14 ms hybrid code search.
Isolation at the OS layer. Scoped task memory (<30 ms Dgraph), cgroup and scheduling-class isolation, and Linux Landlock LSM sandboxing on every agent tool call with immutable audit logs.

Cognitive PC Inc. 5 / 12

Why we can build this

Not a pitch. A codebase and a lab.

Every core claim is backed by public open-source Rust, an operational heterogeneous test bed, and reproducible benchmarks.

34.9k

lines of Rust — hips

Built from scratch by Iason. MIT-licensed, 54 commits, assigned to the company at close. p50 0.58 ms / p95 11 ms on 7.1M articles. Custom inverted index, Block-Max WAND, ggml (Vulkan/Metal/CPU) and CoreML/ANE inference, CLI, VS Code extension and MCP server.

5 nodes

heterogeneous test bed

Three founder-owned nodes (128GB NVIDIA GB10 Grace Blackwell, 2× Apple M4) plus two at the UTH COMNET Lab (Axelera Metis AIPU, 16GB RTX 5080 at 360W). Used for the 15W-vs-360W Pareto and cascade studies, now in preparation.

65k

lines of Rust — spatial runtime

Ilias’s 64,958 net LOC spatial engine (WGSL compute shaders, photonics simulation), a 16-chapter manuscript on stochastic neural networks and queueing, and the PV-CAP-01 KiCad PCB. Private; founder background, not product code.

github.com/jasonthird/high-performance-search-engine 6 / 12

Market & go-to-market

The high-margin local AI workstation.

TAM

$100B+

AI PC and on-prem edge AI appliances by 2030 (MarketsandMarkets, Precedence Research).

SAM

$14.0B

2.8M AI engineers, ML researchers and air-gapped seats in North America and Europe.

SOM

$420M

120k seats — our 36-month serviceable target.

Wedge 1 · AI & spatial developers

Mark I

$3,500 · 37.1% hardware GM

Turnkey 128GB + 15W NPU workstation for 24/7 coding agents and XR studios. $2,200 BOM → $1,300 gross profit. Pays for itself in 4–14 months of displaced cloud API spend.

Wedge 2 · Regulated B2B

Enterprise

$199/mo per seat · 88%+ software GM

Air-gapped workstations for law, defense, health and IP teams forbidden from uploading source code or private data. 100% local knowledge graph, Landlock sandboxing and human-in-the-loop audit logs.

Wedge 3 · Modular upgrade

NPU Kit

$599 · 46.6% GM

A standalone 15W D-IMC M.2 / PCIe NPU plus the cross-platform Cognitive Daemon for existing Linux workstations. Expands software ARR without a full system replacement.

Cognitive PC Inc. 7 / 12

Roadmap & financials

Ship COTS systems today. De-risk silicon on FPGA.

100% of 2027–2029 revenue comes from modular COTS workstations, NPU Upgrade Kits and OS subscriptions. No fab dependency for the revenue in this plan. Revenue figures are targets.

2026 · Gen 0 · Bootstrapped

5 nodes

$0 prior funding raised. hips (34.9k LOC Rust) and a 5-node test bed — Metis NPU, RTX 5080, GB10, 2× M4 — built 100% by the founders.

2027 · Year 1 Target

$1.0M rev

$480k exiting ARR (~800 Pro-equivalent seats). Ship 200 Mark I units ($700k HW + $300k SW, 52.4% blended GM). Port hips to the Metis NPU; RTL on Xilinx FPGA.

2028 · Year 2 Target

$4.2M rev

$2.1M exiting ARR. Ship 800 Mark I/II systems and 350 NPU Kits (53.7% blended GM). Prepare the 16nm/12nm Europractice MPW test shuttle.

2029 · Year 3 Target

$15.2M rev

$6.8M exiting ARR, +$1.72M EBITDA. Ship 2,500 systems and 1,500 NPU cards across US/EU (59.0% blended GM). Tape out custom silicon.

$2.5M Seed Ask = 24 months of runway at zero revenue. Covers ≈$85k/month of OpEx for a 7-person team in Thessaloniki plus $440k of one-off BOM working capital for the first 200 units. Plan revenue and deposits extend runway past Series A.

Cognitive PC Inc. 8 / 12

The co-founders · 50 / 50

Hardware & neural systems meets from-scratch Rust.

Ilias Chrysovergis

Co-Founder & CEO · Hardware systems, heterogeneous compute & runtime

Microsoft Imagine Cup 2016 World Finals — 1st Place, World Citizenship (Seattle, $50k prize, Team AMANDA; cheque presented on stage by Satya Nadella). Co-inventor on granted US Patent #9,965,675 (VR behavioral analysis, 2018).
MSc Imperial College London · MEng ECE AUTH (9.1/10) · PhD candidate, UTH COMNET Lab.
Published in IEEE IoT Magazine (2026), Elsevier C&E: X Reality (2026), IEEE ICMLCN (2025), Springer AACA (2023), ACM SIGGRAPH Posters (2022) and IEEE PIMRC (2020). ~65k net LOC Rust spatial engine and shaders, PV-CAP-01 KiCad PCB, 15W-vs-360W NPU/GPU studies (in preparation).

Iason Malkotsis

Co-Founder & CTO · Rust retrieval engine, Cognitive OS & agent runtime

Creator of hips — 34,922 LOC Rust from scratch, MIT. Hybrid BM25 + dense vector search, CLI, VS Code extension and MCP server. 0.58 ms p50 lexical BM25 across 7.1M Wikipedia articles on an Apple M3; 2.6–14 ms hybrid code search via ggml and CoreML/ANE.
Software Developer, CERTH/ITI & Metatopia. Designed and shipped production human-in-the-loop AI agents across ~7 EU projects; built a Landlock-sandboxed agent prototype with <30 ms memory retrieval; cut PestView query latency ~50%.
Springer Nature co-author (2026) · Kaggle Bronze, Biohub Cell Tracking (268th / 3,947) · BSc Applied Informatics (UoM), MSc AUTH in progress.

github.com/iliachry · github.com/jasonthird 9 / 12

Structure & cap table

A clean Delaware C-Corp. 100% active founders.

Entity

US parent, EU R&D subsidiary

Cognitive PC Inc. (Delaware C-Corp) issues standard YC Post-Money SAFEs and holds 100% of hardware designs, RTL, Cognitive OS and the assigned hips IP.
Cognitive PC Hellas I.K.E. runs the systems and hardware lab in Thessaloniki at a fraction of Bay Area engineering cost. Legacy Metatopia Ε.Ε. projects remain separate.
Both founders full-time at close on 4-year reverse vesting with a 1-year cliff. 3-seat board: two founders, one lead investor or independent.

Post-Seed · $12.5M post-money cap

Aligned ownership

Ilias Chrysovergis, CEO35.0%

Iason Malkotsis, CTO35.0%

Seed investors · $2.5M SAFE20.0%

Employee option pool10.0%

Founders 70% Seed 20% ESOP 10%

Cognitive PC Inc. 10 / 12

The ask · San Francisco & Palo Alto, February 2027

$2,500,000

YC Post-Money SAFE · $12.5M cap · 24 months of runway at zero revenue to ship 200+ Mark I workstations, reach $480k exiting ARR, and demonstrate our custom NPU core on Xilinx FPGA.

Use of proceeds

Cognitive OS v1.0, hips NPU port, daemon — founders + 3 Rust engineers40% · $1.0M

Mark I: full BOM working capital for 200 units, ODM NRE, CE/FCC, demo units28% · $700k

Senior FPGA/RTL architect and Versal boards14% · $350k

Bay Area DevRel, design partners, launch10% · $250k

Delaware spin-out, IP assignment, patents, legal8% · $200k

Live demo

We’re in the Bay Area this February with hips and an early Cognitive OS prototype running live.

Ilias Chrysovergis · CEO [email protected] · iliachry.gr · github.com/iliachry

Iason Malkotsis · CTO [email protected] · malkotsis.com · github.com/jasonthird

Cognitive PC Inc. 11 / 12

Act II · Post-Seed ecosystem expansion · Split-brain spatial computing · $1,499 target ASP ($720 BOM, 52% GM)

The Cognitive XR Headset, tethered to 128GB.

The Cognitive XR Headset paired with the Cognitive PC

Featherweight Spatial Visor (<180g)

Dual 4K Micro-OLED · 90/120 Hz · 4-Cam VIO + Eye & Hand

Moves heavy workstation compute and thermal load off your face onto the Cognitive PC over USB4 (40 Gbps) or Wi-Fi 7.

xrOS Real-Time Microkernel (≤7 ms MTP)

Bare-metal Rust (#![no_std], AArch64) · Hierarchical EDF

≤50 µs sensor interrupt preemption, lock-free SPSC sensor rings, and on-visor Late-Latching ATW reprojection.

Hardware-enforced privacy

seL4-style capability tokens + ε-differential privacy on gaze/pose streams, plus an unpreemptable ≤2 ms emergency passthrough band.

Perception-as-a-Service (15W Metis NPU)

Run-once, share-many VIO SLAM, 5 cm room mesh, 6DoF object pose and encrypted ≤500 ms World Anchors on the 214 TOPS NPU.

3D Spatial Compositor + 70B LLM

Multi-App Spatial Compositor (MASC), OpenXR 1.0, hips 0.58 ms memory and local 35B–70B agents in shared 3D space.

Cognitive PC Inc. · Act II Product Roadmap (Mark I streams to Quest 3, Vision Pro & OpenXR on Day 1) 12 / 12