AiNEKKO AiNEKKO

Agentic
Silicon Factory

From bespoke silicon to programmable manufacturing

Investor Presentation · 2026

Our Purpose

What if every machine could order its own chips?

The semiconductor industry standardized chips to amortize the cost of creating them.

AiNEKKO standardizes the creation process so the chips themselves can be specialized.

Tailored to the workload

Your workload is the spec

Model pipeline, DSP, control loops and latency budget define the silicon. Not a vendor's catalog.

Tailored to the system

Your constraints are the architecture

Power, thermals, form factor, BOM and I/O are design inputs, not integration afterthoughts.

At your cadence

Silicon that keeps up with the roadmap

Products refresh every 6-12 months. Their compute should too, not on a 2-3 year ASIC program.

We sell the outcome

Qualified compute parts with the full software and inference stack, optimized for the end-to-end use case.

Custom-silicon performance, catalog-chip simplicity.

Not a productivity tool for a silicon team

A full solution so you don't need to have one.

The Problem

Every product team is forced to choose between buying a compromise and undertaking expensive custom development

Off-the-shelf silicon

  • Locked before your workload enters the picturePower, thermals, latency, wake-up time, battery life and BOM are fixed upfront. Product engineering goes into adapting to the chips, not differentiating the product.
  • Integration is on youConnecting multiple chips from different ecosystems, model porting and optimization. Innovation is limited.
6–12 mointegration & qualification
2–10×performance loss due to overheating and throttling (customer-provided numbers)

A dedicated ASIC program

  • Fund an entire silicon-development programSilicon team, tools, verification, fab relationships, masks, packaging and bring-up.
  • Carry the entire riskArchitecture choices and production are on one company's shoulders.
  • Wait for working siliconProduct roadmaps inherit the full cost and cadence of a dedicated chip program.
1–2 yrskickoff to working silicon
$20–50M+typical mid-complexity program
Silicon production is still organized and financed as individual engineering programs. As a result, most of the silicon that could optimize real-world systems is never designed

Solution

Agentic Silicon Factory

AiNEKKO is building the next shared production layer: learning, adaptable, AI-native factory capable of creating many different, workload-specific chip products, with each product contributing reusable assets to the next.

The next abstraction layer in semiconductor production. Fabs industrialized wafer fabrication as a service. AiNEKKO industrializes the creation of the chip products themselves.

Economies of scope, not just scale
Catalog & programmable
One design. Many customers. NVIDIA · Qualcomm · Sima
Reuse The same physical chip
Custom ASIC services
One customer. One program. In-house · Broadcom
Reuse Tools & methodology
AiNEKKO · Agentic Silicon Factory
One factory. Different chips. Shared production across workloads
Every new chip improves the next run
Customer propositionCustomers buy chips, not chip-development programs.

Same way as cloud model changed the compute procurement from individual capital-intensive buildouts to shared infrastructure and reserved capacity, customers can procure workload-specific silicon without committing to the full cost, organization, and timeline of developing their own ASIC.

Why Now

Demand for specialized silicon is exploding and the bigger portion of that market is becoming economically feasible

Demand fragments

Physical AI brings increasingly diverse combinations of requirements as robots, drones, sensors and industrial systems each mix their own models, dataflows and hard physical constraints.

Models · Sensors · Latency · Energy · Memory · Physical constraints
The number of economically valuable specialized designs is growing, but few individual products can justify a dedicated ASIC program.

Creation becomes repeatable

Reusable silicon IP has existed for decades, but integration effort was high and expensive

What's changing is the ability to collapse co-design into one end-to-end optimization problem and explore all possible designs not just a single one

Design space exploration · Verification · Physical feedback · Composable silicon blocks
Manufacturing becomes combinatorial rather than a succession of bespoke projects
The inflection

A new economic category becomes possible

Our expert-governed agentic loop crossed the threshold in June 2026: the speed and quality of creating hardware artifacts now justify the customization timelines.

Read the Model-to-Silicon paper ↗
From a real model to hardware artifacts: end-to-end chip design done in 4 weeks by 3 engineers, pushing performance 100× over chips of the same class.

The Economic Frontier

Make previously uneconomic silicon viable

Agentic Silicon Factory expands the economically feasible design space.

A growing share of previously unjustifiable specialized designs can become commercial products.

The market isn't divided into customers who buy catalog chips and customers who commission ASICs. The same customer can move from one to the other as its product matures if the economics make that transition possible.

Until now only high-volume market leaders like Tesla or DJI could do that.

With the factory model the break-even threshold shifts from ASIC-scale volumes toward product-scale volumes and continues moving with accumulated factory reuse.

Moving the break-even point for specialized silicon
Expanded viable market Catalog Custom ASIC Factory Low High New break-even ASIC break-even Product volume Total cost per unit Expanded viable market Catalog Custom ASIC Factory Low High New break-even ASIC break-even Product volume Total cost per unit

Market Opportunity

One production model across different workloads and devices

The newly viable demand emerges in the multiple sectors. Unlike catalogue-chip companies we don’t need to guess how the needs will evolve 2 years ahead and which market will grow to fit it in the architecture, the factory can adjust to the new needs within a single regular cycle.

1 · Physical AI
$103B/yr
edge-AI SoC revenue by 2030, up from $32B in 2024 (21% CAGR), 55% of all SoC revenue
2024 · $32B 2030 · $103B
DefenseDefenseDeliveryDeliveryRoboticsRoboticsConsumer electronicsConsumer electronics

850k robots installed in 2025; 7M+ drones reported in use in Ukraine in 2026; AMD sizes robotics at $200B by 2035.

2 · Reindustrialization of America
~$20B/yr
cost of legacy-chip obsolescence to US long-life industries: last-time buys, forced redesigns, lost revenue (est.)
DoD alone · $2B+/yr ~$20B/yr
DefenseDefenseIndustrialIndustrialHealthcareHealthcare

~475k parts go end-of-life each year; one DoD redesign averages $10M. Equipment lives 20-40 years; its electronics don't. Behavior-preserving replacements as a factory output.

3 · Intelligent Sensing & sensor fusion
$13.5B/yr
4D radar, lidar & sonar, wireless & satellite communications, audio & acoustics, industrial sensing
LiDAR & radar alone · ~$4B ~$20B in 2030
IndustrialIndustrialAutomotiveAutomotiveRoboticsRobotics

Sensor fusion and real-time control run mostly on FPGAs today. We partner with the FPGA: our blocks take the sensor-fusion load alongside it with selective hardening of stable workloads.

Sources: The SHD Group; IFR; AMD; Phoenix Semiconductor; IronGate Capital; Z2Data; DoD DSPO; Altera; Xilinx; Yole Group; MarketsandMarkets

Customer Journey

From off-the-shelf prototype to continuously optimized silicon

Before AiNEKKO

Small volumes

01Prove the use case

  • Off-the-shelf chips, boards and other components
  • Proving the business use case
  • System not optimized

Drone customer case study

Starting system: 4 boards

  • Flight controller
  • Signal processing
  • AI accelerator
  • Host
First AiNEKKO engagement

Initial production volumes

02Optimize the system

  • Customer knows the use case, workload and system requirements
  • Working on Supply chain optimizationBOMBattery lifeProduction complexitySafety
  • AiNEKKO provides the solution now

Drone customer case study

Currently in POC with fabbed out silicon, customization takes weeks as no new tape out is needed

5K units in 2027

Numbers on real workloads

8×performance (comparison with Rockchip)
3×energy efficiency
4→1boards: less complexity, cost and weight
Next design

Scale up volumes

03Become best in class system with vertical integration

  • AiNEKKO delivers highly optimized designs for the next product generations
  • Performance, efficiency and speed of innovation become key

Drone customer case study

Next product generation planned in 2028 with expected volumes over 20K

Newly optimized building blocks go on the next wafer, up to 6 months iteration

Numbers on same workloads with optimized blocks

20×+performance
14×+energy efficiency
No system redesign

Factory “Tour”

Industrializing the process, not the product

01

Automated “production lines”

Model-agnostic, token-optimized system with self-correcting agentic loops.

The factory removes work, rather than merely accelerating it. Each run involves up to a dozen agents in different roles representing months of work of an engineering team.

Production-line dashboard: agentic orchestration graph, program review and tracking, design-space exploration, energy and area analysis, and 3D search at scale

Each design campaign produces two outputs: new silicon and new reusable engineering capability

02

Acquiring best “materials” and expanding the reusable “parts” library

Best-in-class RISC-V manycore developed over 10 years, now fused with our iRAM/MRAM into composable building blocks:

  • RISC-V compute shire
    • 32 RISC-V 64-bit cores with vector + tensor units
    • 4 TOPS INT8 @ 1 GHz, shared SRAM, multicast, HW barriers
    • Runs all dynamic work: attention, convolutions, pre/post
  • iRAM/MRAM compute tile with changeable personalities
    • 4 MiB non-volatile, MAC slices next to the sense amps
    • Weights never move: ~14× less energy/MAC than LPDDR5
    • Per layer: INT8 · INT4 / FP4 · ternary · DSP
  • Scalable, scheduler-agnostic fabric
    • Ultra-scalable mesh NoC (validated up to 4,096 cores)
    • Tiles run descriptor chains from any core or host
    • Hardware credits signal completion: no polling

Same parts, different SoCs (TSMC 16 nm, design estimates): GenAI-first 1 shire + 16 MRAM tiles ≈ 91 mm² · Physical AI 4 shires + 8 tiles ≈ 100 mm² · Generalist 4 shires + 16 tiles ≈ 115 mm²

Cores extensively tested and taped out in the 1,088-core Esperanto ET-SoC-1, since validated by the open-source community at the OpenHW Foundation.

03

Training datasets

Acquired an enormous digital footprint of Esperanto Technologies corporate and technology data (internal communications, operational & corporate records, the complete R&D track record meticulously preserved across design, validation, manufacturing and customer PoCs) going all the way back to 2015.

Used as the seed dataset for our self-correcting agentic-loop training approach, and it is already providing us with major engineering breakthroughs. Since the initial training and pipeline setup, we have engaged in folding our customers’ agentic trajectories into the same dataset to improve its accuracy and generative characteristics.

04

Availability of supply chain

All critical silicon IP blocks are owned by us, open source and transparently audited, and already in the training data of the agentic system no 3rd-party licensing risk or price hikes.

Manufacturing capability: optimizing for cost, waiting times and energy efficiency of nodes, currently taping out at TSMC 12/16nm and closely collaborating with TSMC’s R&D team at the cell and technology level, aligning with their roadmap. We do not depend on high end processes and can tape out at other fabs/geographies.

Exploded MoSIC package composed of reusable chiplets

Advanced packaging capability: we currently partner with Deca Technologies for advanced fan-out wafer-level packaging (FOWLP) and heterogeneous integration, optimizing for time-to-product, capabilities and cost. We are working on diversifying our packaging suppliers over time.

Business Model

We accumulate designs into regular tape-outs, keeping the factory utilized with high-volume channel partners

1

Direct sales

Sell the outcome, then the design-in follows

We approach companies going into volume (3,000+ units), knowing the workload and constraints and optimizing the production of the system — silicon should not be treated differently. We aim to build long-term relationships with regular cycles of silicon updates.

Current pipeline has drones, warehouse robotics and LiDAR customers at different stages of evaluation

2

Channel partners

One design-in, many products

Platform owners aggregate the customers’ needs for semi-custom solutions, carrying them into many products — cheaper customization, higher volumes.

Developer platforms (e.g. Raspberry Pi, BeagleBoard, Lattice Semiconductor)Modules become readily available AI options on maker-to-product platforms.
Autonomy kit builders (e.g. Auterion, Airvolute)One kit design-in reaches many airframe OEMs and integrators.

Current pipeline justifies the cadence

2 tape-outs in 2027 → 4 regular tape-outs per year · no custom tape-outs, designs share the shuttle
Weighted pipelineRaw pipelineRegular tape-out · $2M mask setVolumes, values and dates are customer indications, not bookings

Share the cost of each run

$2M

Tape-out mask cost — shared wafer for all designs, so one mask set is amortized across every program.

Get custom package

$10–100K

Custom packaging

The Team

Built to win this race with experience and execution

We have built, bought, sold and operated computing systems end to end, in both traditional and AI silicon.

  • Tanya Dadasheva, Co-founder & CEO: Semiconductor and open-source investor at Almaz Capital turned operator. Brings go-to-market depth across traditional and AI silicon
  • Roman Shaposhnik, Co-founder & CTO: Former CTO & co-founder at ZEDEDA; previously Pivotal, Cloudera, ASF. Open-source veteran and systems architect. Brings the community playbook that makes open silicon viable
  • Doug Smith, Co-founder & CSO: Former CEO at Veevx; 18 years at Broadcom leading central engineering. Brings broad silicon expertise, MRAM/iRAM and a physical-design team with dozens of tape-outs
  • Dan Stolyarov, COO: Silicon M&A and investment background. Runs operations, program execution and delivery across the silicon roadmap
  • Allen Rush, Chief Architect: Former Senior Fellow and Chief Architect, Imaging & AI at AMD. Founder of Sierra Imaging and PCC; earlier Intel. Leads the architecture roadmap
  • Gianluca Guida, Founding Engineer & VP of Agentic Engineering: Former Sr. Engineer at RIVOS; previously Apple, XenSource, VMware

Team

Previously building Esperanto, Tenstorrent, AMD GPUs, central engineering Broadcom, edge market sales/bizdev

Traction

From conventional engineering to the operating factory

2015–2025

Esperanto Technologies builds manycore RISC-V energy-efficient silicon fabric: 10 years, $200M+, tape-out-proven design

Nov 2025

AiNEKKO acquires Esperanto IP. Announces open-source strategy and platform development

Jan 2026

Merger with Veevx adds MRAM/iRAM and a seasoned physical-design team

Apr 2026

Compute and memory tiles tape out and are contributed to the OpenHW Foundation

Jul 2026

Open hackathon with Hugging Face showed that any model can be mapped to our substrate and optimized agentically

Nov 2026

Factory open for business. POCs with clients

Next

Industrialization of production

The Opportunity

Industrializing the creation of
the next thousand chips

We are raising $20M to enable it

nekko.ai
Website
tanya@nekko.ai
Contact

© 2026 Ainekko, Co.

1 / 15