A new model architecture. Built to compete anywhere.
The incumbents scaled up. We started from first principles. The architecture runs where no other model can. This round funds the proof.
Backed & supported by
Startup ProgramMicrosoftFounders HubInceptionNational Compute GrantNatWestAcceleratorAppwriteStartup Program
ANT
02 / 14
What ANT Is
Same capability as ChatGPT. 60× fewer parameters.
Not compression. A completely new architecture trained from scratch. Validated across five modalities.
LanguageVisionDiffusionDNAProtein
60×
fewer parameters
Today: 7B-equivalent running on an Apple Watch.
This raise: 70B+ on NVIDIA supercomputer. Scaling curve. OEM conversations.
ANT
03 / 14
The Problem
Running AI is becoming unaffordable and inaccessible.
+172%
DRAM YoY
Sold out
HBM 2026
$1B+/yr
To serve one frontier model
1 GB HBM requires 4× the fab capacity of standard DRAM. Not cyclical. Structural. AI will consume 20% of all global DRAM wafer capacity in 2026.
DRAM price index (YoY)
Q1 '25
+28%
Q2 '25
+55%
Q3 '25
+172%
Q1 '26
+55%*
*TrendForce forecast · Jan 2026
ANT
04 / 14
Why Now
Three forces converging. A structural shift.
The memory crisis, budget pressure, and regulation are not temporary. They create a permanent market for architecturally efficient AI.
01
Memory crisis. HBM sold out through 2026. Every AI model needs memory the market can't supply.
02
Budgets breaking. Inference costs make most AI agents commercially unviable.
03
Regulation arriving. EU AI Act, DORA. Regulated industries need on-premise AI.
Google · March 2026
Published TurboQuant."build less memory, not more." Memory stocks fell the same day.
TurboQuant makes HBM more efficient for cloud. ANT built models that never needed HBM. Different starting point. Different buyer.
Key fact
AI will consume 20% of all global DRAM wafer capacity in 2026. For the models our customers will deploy, ANT runs on hardware they already own.
ANT
05 / 14
Current Solutions
They shrink models. We built a new one.
Everyone else
Quantization & pruning
Take a big model, make it smaller. 2–4× reduction. Quality degrades. Hard ceiling. Still dependent on someone else's architecture.
0×4×
ANT
New architecture from scratch
Up to 60× smaller. Quality preserved. No ceiling. Natively designed for constrained hardware, not an afterthought.
0×60×
ANT did not build a better compression tool. We built a new kind of model.
ANT
06 / 14
Competitive Landscape
Everyone optimises existing architectures. We built a new one.
Cloud onlyAny hardware
Small footprintLarge footprint
TurboQuant
MoE / DeepSeek
Quantization
llama.cpp
ANT
Quantization
Small models
ANT
Approach
Shrink existing
Train small from cloud arch
New architecture
Reduction
2–4×
Natively small but limited
Up to 60×
Offline
Limited
Partial
Yes
Quality
Degrades
Good but ceiling
Preserved
Hardware
GPU needed
GPU needed
Designed for constrained hardware
Google · March 2026
Published TurboQuant. "Build less memory, not more." Memory stocks fell the same day. The compression thesis is validated. The regulated and offline market remains unaddressed by anyone else.
ANT
07 / 14
The Approach
Proven at 1.5B and 7B. Now scaling commercially.
1.5B on Apple Watch. Tiny Stories validated on dual-core CPU. No AI silicon, no internet. A fundamentally new architecture, not compression.
7B on NVIDIA supercomputer. Validated with limited data using national compute grant. The architecture holds at scale.
Five modalities. Language, vision, diffusion, DNA genomics, protein. Each trained from scratch, each benchmarked against models many times larger.
Hardware validated
GPUCPUFPGAASICWearable
Validated
1.5B on Apple Watch
Tiny Stories. Dual-core CPU. No Neural Engine. No dedicated AI silicon.
Validated
7B on NVIDIA supercomputer
Limited data. Architecture confirmed at scale.
This fundraise
1.5B commercial, 7B reasoning, 70B proof
Compete with Qwen at 1.5B. Add reasoning at 7B. Secure GPU access for 70B.
ANT
08 / 14
What This Unlocks
OEM is the entry point. Big ticket, high margin.
01
OEM Licensing
Architecture at chip/firmware level. Royalty per unit shipped. ARM, Dolby, MPEG-LA precedent.
Big ticket
02
Offline
No internet required. Field engineers, remote infrastructure, logistics.
2 customers live
03
Privacy-first
Data never leaves. Pharma, legal, defence. On-premise, fully auditable.
Active traction
04
Agentic AI
Low-footprint adapters. E-commerce, chatbots. Unit economics that work.
Experimenting
ANT
09 / 14
Traction
Six months. Revenue, grant, IP.
3
Customers
£18K
ARR
~£50K
National Compute Grant
All traction achieved in six months from platform launch, with zero external funding. Three paying customers across two verticals. Patents filed (pending). Data partnership signed.
Nov '24
Platform launch
Dec '24
First customer · £18K ARR
Jan '25
National compute grant · supercomputer access confirmed
Feb '25
Customer 2
Mar '25
Customer 3 · patents · data partnership
2026
Supercomputer benchmarking · Series A
ANT
10 / 14
Business Model
Three revenue tiers. One architectural core.
01
Now
Platform & services
Fixed annual fee per deployment. Enterprise agentic platform (Everyday Series). Long term: customer data feeds back into building their own SLM, which uses very little compute and is cheap to run.
Recurring ARR
3 customers · live
02
Mid-term · high margin
OEM & architecture licensing
Architecture licensed at chip or firmware level to device OEMs and hardware manufacturers. Royalty per unit shipped. Scales with customer shipment volume, not headcount. FPGA and circuit-level validation complete.
Per unit royalty
Scales without headcount
03
Ongoing
Research & grants
National compute grants, data partnerships. Non-dilutive. Signals credibility to co-investors.
Non-dilutive
~£50K confirmed
Platform ARR now. Licensing royalties at hardware scale. One core. Three revenue lines.
ANT
11 / 14
Team
The founding team.and who has committed to join.
Founder & CEO
Dr Gaurav Gandhi
PhD nonlinear systems & chaos theory. Five years chip design at STMicroelectronics and Cadence. Prior company built and exited. RAEng (LiF) Fellow. Architecture traces directly to doctoral research, a decade of prior work.
CCO · Joining
Founded, scaled, and exited tech businesses. Military career culminating in operational MBE. Deep tech ventures, public sector procurement, defence contracting.
Engineering · Joining
Previously at Automattic (WordPress/WooCommerce) and Deel. Built and shipped production systems at global scale.
Research · Joining
Cambridge PhD, mathematics. Mathematical formalisation, IP documentation, and publication strategy.
Advisors
Active coverage across pharma, chip design, e-commerce, and venture.
Raise closes the team. Commitments in place
ANT
12 / 14
Use of Funds
Build the models. Find the market.
Build
01
1.5B → Commercial. Compete with Qwen at a fraction of the size. Production-ready release.
02
7B → Reasoning. Advanced capabilities. Enterprise-grade. Test innovations and emergent behaviors.
03
70B → Prove scale. Secure GPU access. Publishable scaling curve. The Series A data point.
E-commerce / chatbots. Low-footprint adapters. Experiments in progress.
03
OEM. Licensing conversations with hardware partners. Big ticket opportunities.
Double down on whichever vertical shows the strongest signal.
ANT
13 / 14
12 Months
Build. Explore. Double down.
01
Ship 1.5B: production release, compete with Qwen at a fraction of the size
02
7B with reasoning: enterprise-grade capabilities, test emergent behaviors
03
Secure GPU access for 70B: publishable scaling curve, the Series A data point
04
Run vertical experiments: enterprise, e-commerce, OEM. Find the winner.
05
Series A: scaling data + market clarity + IP
ANT
14 / 14
The Ask
Seed Round
2026
We have revenue, a grant, patents pending, and a team with the specific background this problem requires. The architecture works at small scale across five modalities. The raise funds the large-scale benchmarking that shows how far it goes.
We are not asking you to fund an offline AI company. We are asking you to fund the evidence that a new architecture, validated across five modalities, already running on hardware the incumbents cannot reach, belongs in the same conversation as the best AI in the world.