HORIZON 1NOW
NOV 2026 – MAR 2027
Measure
- NOV–DECTier 0 · Open energy ledger
- DEC–JANTier 1 · Reference and baselines
- JAN–MARTier 2 · One layer on FPGA
◆ GATE
One audited energy-per-token result the field can reproduce.
01THE MEMORY WALL
A petaflop on the desk.
Twenty tokens a second.
NVIDIA's DGX Spark puts a petaflop of FP4 on a desk. For one user it decodes Llama 3.1 8B at 20.5 tokens a second, because every token streams the model's weights from memory.
FIG. 1 · THE WALL IN A SHIPPING PRODUCT
NVIDIA DGX SPARK · 1 PFLOP FP4 · 273 GB/S · LLAMA 3.1 8B, FP8
18×
SOURCE · LMSYS, OCT 2025 · SGLANG
more output once 32 streams share each pass over the weights. Batching hides the wall in a datacenter. A device serving one user can't batch its way out.
FIG. 2 · WHERE ONE TOKEN'S BYTES GO
QWEN3.5-4B · BATCH 1 · INT4 WEIGHTS · FP32 STATE · 1K CONTEXT
Weights2.1 GB94% OF BYTES
Recurrent state101 MB4.5% OF BYTES
Attention KV cache34 MB1.5% OF BYTES
| THE AMDAHL TABLE | CEILING FROM STATE ON CHIP |
|---|---|
| One state-update kernel | 13–24× |
| Whole model, one stream | 1.05× |
| Whole model, 16 streams | 1.8× |
A kernel can show 13–24× while the whole model moves 1.05×. We publish the whole model.
Their answer is more batch and more capacity.
Ours is fewer bytes per token.
02THE GAP
They're shrinking datacenters onto desks. We're making intelligence a component.
ON A DESK
DGX Spark · GB10
Built forAI developers prototyping models
Sold on1 petaFLOP FP4 · 128 GB
Price$4,699 list
BESIDE A DESK
Pro Max GB300
Built forEnterprise AI teams
Sold on20 petaFLOPS FP4 · 748 GB
PriceTower-class
IN A ROBOT
Jetson
Built forRobotics OEMs
Sold onTOPS, in 7–130 W modules
Price$249–3,499 kits at launch
INSIDE THE PRODUCT
Babylon
Built forDevice makers shipping machines
Sold onJoules per token · idle watts
PriceModule-class
THE NUMBER NOBODY LEADS WITH
None of them headlines joules per token or idle watts. Reviewers measured a DGX Spark at 25–45 W sitting idle. For a device that is always on, those are the numbers that decide the product.
NOT OUR LANE EITHER
Crowd-scale serving in the datacenter (Groq, Cerebras), or one model frozen into silicon (Taalas, Lamb Labs). Our architecture stays programmable across the model family.
SOURCES · NVIDIA, DELL, LMSYS, TOM'S HARDWARE, SERVETHEHOME
03WHY NOW
Three things changed at once.
In the past year the models, the buyers and the price of memory all moved. Each one points at the same product.
- 01 · THE MODELS
Models changed shape.
Qwen3.5, Nemotron 3 and Kimi Linear ship recurrent layers next to attention, and those layers keep a fixed-size state instead of a growing KV cache. That is the model class our hardware is native to, so our model bet and our hardware bet are the same bet.
- 02 · THE BUYERS
Local AI found its buyers.
NVIDIA and Dell now sell AI computers for the desk. Running AI off the cloud is proven demand. The always-on device inside a product is still unserved.
- 03 · THE MEMORY
Memory became the bill.
DRAM contract prices nearly doubled in the first quarter of 2026 alone, and NVIDIA raised the DGX Spark's price by $700, citing memory supply. Every byte a token moves now costs energy and bill of materials.
SOURCES · QWEN, NVIDIA, MOONSHOT AI, TRENDFORCE
04THE ROADMAP
Measure first.
Shrink the bytes.
Then change the substrate.
HORIZON 2
2027
Shrink the bytes
Move fewer bytes per token, and keep state resident across many always-on streams. First eval kits and paid pilots.
◆ GATE
≥10× vs datacenter GPUs and 2–3× vs Jetson, measured. One paid pilot, two letters of intent.
HORIZON 3
2028+
Change the substrate
Memory and compute as one physical device, built with research labs. It opens workloads no NPU can run.
◆ GATE
One lab partnership signed. A taped-out cell characterized.
EVERY NUMBER WE PUBLISH STATES
Scope · board and baseline · model and checkpoint · precision · batch and context · quality against the reference · energy at the wall · latency · idle power and wake time
05WHAT WE SELL
A socket and an API swap.
A compute module and runtime that give a device maker an always-on language model inside their power budget, for machines, robots and appliances that run unattended for years.
- 01 · PILOTS
Paid pilots
€20–50k NRE: we port your workload, then deliver a measured report and a working prototype.
FIRST · 2027 - 02 · EVAL KITS
DK-1
An eval kit on AMD's K26 production SOM. Measure joules per token on your own bench.
2027 - 03 · MODULES
B1
A compute module on a Kria-class SOM, designed into your product as a socket.
NEXT - 04 · SILICON AND IP
Silicon and IP
The architecture as silicon or licensed IP.
LATER
SAME CODE AS THE CLOUD
An OpenAI-compatible local endpoint, Hugging Face and GGUF checkpoints, LoRA as weight downloads. The app developer never sees an FPGA.
SUPPLY YOU CAN SIGN FOR
Silicon supply is AMD's. The K26 production SOM carries ten-year availability, so we ship architecture and runtime, not a factory.
ESCROWED ARCHITECTURE
The architecture is held in escrow, so your design-in is protected for the life of your product.
06ABOUT · WORK WITH US
Help lay the next tier.
Babylon is an early-stage computer architecture company building always-on inference for recurrent and hybrid AI models. We come from systems software: model representation, compilers, runtimes and benchmarking, the layer where architecture meets the model. We ship architecture and runtime, and partner for silicon and new materials.
- A hardware architect, as advisor or co-founderNOW
- Design partners with an always-on, on-device workloadNOW
- Programmes and investors backing deep-tech hardwareNOW
- Research labs building memory that computesOPEN