Frequently Asked Questions

Questions grouped by who is asking: investors, technical evaluators, and everyone else.

For InvestorsFor Technical EvaluatorsGeneral

For Investors

What problem does Layer1Labs solve, in business terms?

AI training runs on rented or owned GPU time, and a large share of that time is wasted: expensive accelerators sit idle while checkpoints write to storage and data loads. Layer1Labs' ChronoHive coordinates that I/O work deterministically, recovering wall-clock time — which converts directly into lower cost per trained model and per token served. The need is broader — deterministic scheduling across software and hardware systems — but training is the beachhead, where GPU time costs most today.

How does a training speedup turn into money saved — what is the token economics?

Token economics is cost per token: every AI output is produced by GPU-hours, so cost per token falls when the same GPUs finish more work per hour. ChronoHive's measured speedups mean a training run completes sooner on the same hardware — fewer GPU-hours purchased for the same result, or more tokens produced from the same fleet.

What has Layer1Labs actually proven so far?

On real data, ChronoHive delivered a measured +9.8% end-to-end training speedup in a preregistered A/B trial — the test design was fixed in writing before any run. In the checkpoint regime where production clusters hurt most — GPU-bound training over slow, shared storage — measured wall-clock gains reached +24.1%. Every published figure carries its measurement label.

What could the savings look like at cluster scale?

Only as a labeled illustration — not a measurement: Layer1Labs' investor materials model roughly $2.7M per year for one 1,024-H100 cluster IF a 5% end-to-end gain materializes, using published cloud GPU pricing. No cluster-scale measurement exists yet; producing one is precisely what the current raise funds. Treat the scenario as arithmetic, not a result.

What is the seed round, and what does it buy?

Layer1Labs is raising $1–1.5M on a SAFE, seeking a lead investor, with preseed complete. The round buys cluster-scale proof: multi-node validation from 8 to 32 GPUs on shared storage, a joint design-partner benchmark report, and a pilot-ready release with a Series A data room by month 12. Investor inquiries go to investors@layer1labs.ai.

How does Layer1Labs make money?

ChronoHive is licensed as software into the clusters customers already own — no hardware purchase required — and both ChronoHive and ChronoFabric are offered for commercial licensing by written agreement. Software licensing into existing infrastructure is the near-term revenue path; the ChronoFabric hardware tier deepens the same architecture for partners who want enforcement in silicon.

What is Layer1Labs' intellectual property position?

Layer1Labs has 10 patents pending, covering the stack before it is public — the admission control kernel, asynchronous checkpointing, and the ChronoFabric effector fabric. The products are proprietary technology of Layer1Labs Silicon Inc., licensed — never sold — by written agreement, so the implementation stays closed while customers and partners get its measured outcomes.

Investor inquiries: investors@layer1labs.ai

For Technical Evaluators

How does ChronoHive work, at a technical level?

Work is written as programs in Lingua Franca — deterministic, timed reactors (self-contained components that react to events) — and compiled by chronoc into a single constraint specification (.cspec). The ChronoHive engine executes the specification, enforcing the program's compiled policy for prefetch, checkpointing, and dispatch deterministically, instead of letting a heuristic scheduler guess at runtime.

What does "deterministic" mean here — what exactly is guaranteed?

The same program with the same inputs produces the same decisions, bit-for-bit, on every run — so any run can be replayed exactly for debugging, regression testing, or audit. In orchestrator campaigns, 92 of 92 pipeline runs reproduced the reference outputs exactly, with zero duplicate writes — determinism is measured, not just designed.

What is the headline benchmark, and how was it measured?

The headline is a measured +9.8% end-to-end training speedup on real data: a preregistered A/B (ResNet-18 on CIFAR-10, six interleaved 200-step trials, on two GPUs), with the design fixed before any run. Checkpoint integrity held alongside: 400 of 400 checkpoints readback-verified, and final training states bit-identical in 20 of 20 trial groups.

How do the measured results vary by storage regime and workload?

Regime decides everything. GPU-bound training over slow, shared storage (~100 MB/s) measured +24.1% and +21.7% wall-clock by GPU; on fast local NVMe the same class of workload measured +5.9% / +4.9% with 94 MB checkpoints and +7.3% / +6.5% with ~1 GB checkpoints (RTX 4070 SUPER / RTX 5060 Ti); two-GPU distributed training measured +1.7% end-to-end. Figures come from separate preregistered campaigns and are never stacked or merged.

Where does ChronoHive not help — what are the measured limits?

We publish the misses. CPU-bound workloads on fast NVMe still trail synchronous checkpointing by 6–14% in our measurements. Very large checkpoints (~4 GB) written at short intervals (~3.5 s) are the hardest case we have measured: the best configuration — a coalescing writer whose memory stays bounded at 3× checkpoint size — still trails synchronous checkpointing by about 32%, because encoding the checkpoint takes ~10 s against a ~3.5 s interval; the limit is encode speed, not memory. Admission gating measured +9.85% on an Amazon EC2 g5.xlarge instance (NVIDIA A10G), +7.95% on our RTX 4070 SUPER, and +5.54% on our RTX 5060 Ti (22 completed preregistered trials) — positive on every platform, but below our own 10% confirmation bar, so we report it as an observation, not a proof.

What is the validated scope — and what has not been validated yet?

Every measurement behind these answers is single-node, on at most 2 GPUs — no multi-node or cluster measurement exists yet. A scalability model (a simulation, not a result) suggests a cross-regime median of +12.5%, but the funded next step is the real test: multi-node validation from 8 to 32 GPUs on shared storage.

How does ChronoHive integrate with our existing infrastructure?

You keep your infrastructure: ChronoHive runs on the clusters you already own, with no hardware purchase. Work is authored as Lingua Franca programs; their effectors — the connection points between the program and the outside world — bind to your storage and compute through software callables, and the technology is licensed by written agreement, with an evaluation path available. Start via licensing@layer1labs.ai, or info@layer1labs.ai for technical questions.

What is ChronoFabric's technical status — and does it share ChronoHive's guarantees?

ChronoFabric runs the identical programs and constraint specifications as ChronoHive; only the effector binding changes — to FPGA and hardware endpoints instead of software callables. It is complete and ready for beta customers, and its admission, sequencing, and determinism guarantees carry over unchanged. ChronoFabric is available now as soft IP — a licensable hardware IP block — targeting Kria FPGA platforms. Sustained, not just sampled: over a full 24-hour run on a Kria KV260 board — 864 million timing cycles — ChronoFabric's real-time software stack (PREEMPT_RT kernel with isolated CPUs) held worst-case scheduling latency to 22 microseconds on the isolated core, with 99.99% of cycles at or under 12 microseconds. The 24-hour run was completed for all three software configurations on the board: the full hardware profile passed (worst case 22 microseconds), the real-time software configuration passed on its own (worst case 33 microseconds, 99.99% at or under 15 microseconds), and the stock configuration, measured for contrast, reached a worst case of 904 microseconds (99.99% at or under 25 microseconds). The timing mechanism was also validated at the tighter budgets faster links would impose — up to the 620-nanosecond cadence a 1.6-terabit link would require, holding the schedule exactly at every level; that validates the timing mechanism at those budgets, not data transport at those line rates. Saturation-proof: with all four CPU cores held at 100% load, ChronoFabric's scheduled departures still landed on the programmed tick (median and p99 exact), while CPU-scheduled departures on the same boards degraded to millisecond-scale jitter. The hardware scheduler issues no per-packet work to the CPU at all. Embedded chiplets and merchant silicon are future offerings on the staged hardware roadmap.

Partnerships & licensing: licensing@layer1labs.ai · Technical & beta questions: info@layer1labs.ai

General

What is Layer1Labs?

Layer1Labs Silicon Inc. is a Birmingham, Michigan company building deterministic orchestration technology for heterogeneous compute — the mix of CPUs, GPUs, and accelerators modern AI runs on. Its thesis: AI scales as physical infrastructure, so performance should come from deterministic execution scheduling, not from buying more hardware to cover coordination waste.

What products does Layer1Labs offer?

Two tiers of one architecture. ChronoHive is the software tier: a deterministic execution engine for AI and data infrastructure, licensed into the clusters customers already own. ChronoFabric is the hardware tier: the same programs and constraint specifications, enforced on FPGA and hardware endpoints. ChronoHive is available for licensing; ChronoFabric is complete and ready for beta customers. ChronoFabric is available now as soft IP — a licensable hardware IP block — targeting Kria FPGA platforms; embedded chiplets and merchant silicon are future offerings on the staged hardware roadmap.

Who is Layer1Labs' technology for?

It is for organizations that train AI models or run large data pipelines on expensive accelerators — and pay for GPU time that coordination waste burns. Infrastructure, platform, and AI teams use it to recover throughput from the clusters they already own; hardware builders can take the same architecture into silicon through the ChronoFabric beta.

Is this only for AI training?

No. Training is where Layer1Labs' measured results are today, but the technology is a general scheduling paradigm: systems are written as deterministic reactors in Lingua Franca — timed components that react to events — and the same programs schedule work in software with ChronoHive or in hardware with ChronoFabric. Any system of timed, interacting parts — industrial control, robotics, infrastructure automation — fits the same model.

How is Layer1Labs different from a conventional scheduler?

Picture a big city's traffic lights: today each light runs its own fixed timer, blind to the grid — so traffic stalls even when the road ahead is empty. A coordinated grid could time every light against actual traffic, making green waves instead of gridlock. Layer1Labs does that kind of coordination for compute: instead of heuristic schedulers that guess, the optimization policy is compiled into the program and enforced deterministically — same inputs, same decisions, bit-for-bit, every run replayable.

How do I get started with Layer1Labs?

Start with a conversation. For partnerships and licensing, email licensing@layer1labs.ai; for technical and beta questions, info@layer1labs.ai; for investor inquiries, investors@layer1labs.ai. Pilots are isolated and software-first — no hardware purchase — with the validation methodology, arms, and success criteria fixed in writing before any run, and a joint benchmark report as the deliverable.

General contact: info@layer1labs.ai