Accepted agent tasks per kWh.
Accepted tasks (or subagent steps) per kilowatt-hour at full operating power, counting the whole serving stack, not chip-only energy.
Project Vela / The defining objective
One launch. A 100,000-GPU data center. Powered by sunlight.
Infinite Sunshine. Effortless Intelligence.
Data in, intelligence out. Vela is a solar-powered agent firehose: high-bandwidth inference for many concurrent low-to-mid weight models, measured in accepted agent work per kilowatt-hour, per dollar and per launched kilogram.
Our benchmarkAccuracy gate passed at small-model scale; larger models in progress · Updated Nov 2026
COMPUTE / ARCHITECTURERESEARCH NOTES — 01
From photons
to useful inference.CONCEPTUAL
ILLUSTRATION
to useful inference.CONCEPTUAL
ILLUSTRATION
Primary focus and exploratory work
Orbital compute systemsNovel inference architecturesBenchmark / Updated Nov 2026
Our benchmark: tokens per solar watt
In orbit, power is the one thing every system gets equally: about 1,361 watts of sunlight per square metre. So we measure what matters: how much useful AI work each watt of sunlight delivers. We quote it for an 8B-class language model, and convert it to the GPU everyone knows.
| Generation | Status | What changes | Tokens per solar watt | vs. an NVIDIA H100, per watt |
|---|---|---|---|---|
| V1 | Building now. Measured on the bench | Off-the-shelf low-power AI chips on a solar panel | 14–18 | On par |
| V2 | In development with chip suppliers | A custom chip designed for this one job | 200–370 (target) | ~10–18× |
| V3 | In design | The panel itself becomes the computer | 650–1,300 (target) | ~30–65× |
V1 proves the system. V2 and V3 multiply it. Each step turns more of the sunlight into answers.
How we get there, and why we believe it → · Read the paper →
H100 basis: MLPerf Inference, 8 × H100 on Llama 2 70B, at server power, scaled to 8B class. Our watts are what the solar cells deliver. V2 and V3 are design targets.
Design target / One launch
What one launch is designed to do
~5 billion tokens per second on a 7B-class model, in sunlight
≈ 100,000 H100 GPUs of delivered AI work
~6 MW of solar power, where the same work on the ground draws 100+ MW of grid power
80 satellites per 100-tonne Starship launch
Mature chip processes: no EUV, no photonics, no new materials
Third-generation design target. Not yet built or flown. Why we believe it → · Read the paper →
The building block: one compute panel
~2.9 kW of compute per panel (nameplate: new cells at standard test temperature)
4.35 × 2.40 m, made of tiles that each carry their own solar cell and compute
Spare tiles built in, so a failed tile is switched out, not repaired
Figures in the Benchmark and Launch sections for V2 and V3 are design targets. V1 and the Bench results section are measured.
Watch the overview
Project Vela in a few minutes.
Φ / What we measure
Accepted work is
the measure of success.
We score completed agent tasks that pass a predeclared quality bar, not raw tokens and not efficiency slogans. Tokens are not completed tasks.
Dollars per accepted task.
Cost per accepted task or step, including failed attempts, retries and allocated capital cost.
Accepted steps per second.
Accepted steps per second per panel (or per kW), held to the same quality bar under high concurrency.
Useful tokens per joule.
Reported for the low-to-mid models in our test suite only. We do not claim or compare frontier-model tokens per joule.
“Dollar-Kg” expresses a joint optimization of the accepted work above against cost and launched mass. Per kWh is the energy truth, per kg is the launch truth, per dollar is capital recovery. We report all three on the same task suite rather than collapsing them into one score.
Why Vela / The power gap
Demand is outrunning
the power to serve it.
Trend extrapolations [CALC], not forecasts. There is no public token series for subagent inference.
A gap of roughly 25×.
In a base case five years out, inference token demand grows about 75× from today while AI data-centre power grows only about 3× (IEA-shaped trend). Useful tokens per joule must rise roughly 25× or demand cannot clear.
Volume alone won’t close it.
On Earth there is not enough power for a pure volume ramp. In orbit, sunlight is a level field of about 1,361 W/m² for everyone, so the win is useful intelligence per dollar, and per kilogram and watt as physics.
Specialized subagent inference.
Vela targets high-concurrency, low-to-mid weight “sniper” subagents. For how our efficiency compares with a GPU, see our benchmark table.
Vela is a capacity multiplier inside the ~25× gap, not a claim to create the whole 75× market.
Design targets / Not shipped
What one panel
is designed to do.
Per-panel design targets [CALC] for the subagent regime. Accuracy gate passed at small-model scale; larger models in progress.
| Meter | Design target |
|---|---|
| Panel envelope | ~40 kg · ~2.9 kW of compute per panel (nameplate) |
| Accepted subagent steps / kWh | ~1×10⁵ – 5×10⁵ |
| Accepted steps / s / panel | ~50 – 250 |
| Tokens / s / kg diagnostic | ~2×10³ – 5×10³ |
| Model sweet spot | Built for 3–8B models today; larger models on later generations |
Targets will be revised as testing replaces assumptions. Measured results are in Bench results.
01 / Orbital compute research
Project Vela.
From photons to useful work.
One modular architecture, from a single solar-compute tile to a full spacecraft.
01
Physical systems
Project Vela
Concept & engineering studies
Compute that lives on the solar array.
Modular from tile to spacecraft.
Every compute tile pairs its own solar power with its own inference silicon. Tiles combine into groups, groups into panels, and panels into wings around a central hub. The same building blocks repeat at every scale, so the design grows by adding panels rather than redesigning the spacecraft.
- TileSolar power and inference silicon on one surface.
- GroupNeighboring tiles that share local coordination and storage.
- PanelA repeatable, manufacturable unit with its own gateway.
- WingsPanels that fold for launch and deploy in orbit.
- HubDispatches work and links the spacecraft to the ground.
Agent firehose, not frontier rival. Vela serves many concurrent low-to-mid weight models (specialist and subagent models; built for 3–8B models today; larger models on later generations). It is not trying to outrun frontier GPUs on chat quality. Product Gen‑1 proves the platform and does in-orbit data work; LLM-scale token service is product Gen‑2. (Separately, the panel hardware design is in its third iteration.)
Designed for agents, not chatbots. Vela prioritizes reliable autonomous productivity over immediate conversational response. Chat-style use is still supported; it just isn’t what the system is optimized for.
Built to keep working. The architecture is designed so that no single tile is critical and the spacecraft keeps delivering useful work as individual parts age or fail.
Generations. We build on mature, high-volume chip processes that avoid the EUV bottleneck and its cost premium. Vela targets inference, not training; training is left to the existing players.
Why specifics are not published yet
Panel designs are now in their third generation of iteration, with more to follow. High-level panel size and nameplate power are on this page; chip, tile and method details stay private until intellectual-property filings and hardware validation. We share architecture and test data with researchers, partners and investors under NDA: info@project-vela.com.
Current objective
Vela-Scout 83.
A rideshare initiative.
Our first spacecraft objective: a proposed demonstrator built from the same panel architecture and sized for a SpaceX Falcon 9 rideshare launch. Not yet committed.
Current objective / Falcon 9 rideshare
Vela-Scout 83
Demonstrate solar-powered inference in orbit with flight hardware from the same architecture that later spacecraft will use.
What it should show
Useful output in orbit
Power, thermal behavior, data links and useful inference measured on a real spacecraft, not just on paper.
Future work
Vela-Optima & Vela-Max
Larger multi-panel spacecraft concepts that build on Vela-Scout 83 results. Specifications will follow validation.
Spacecraft-level projections will be published as they are baselined. Figures elsewhere on this site are design targets. The Bench results section is measured.
02Novel
Exploratory
Novel
Architectures
Early exploration · future work
New ways to run
inference efficiently.
Alongside Vela, we are exploring new system and software architectures that could make inference more efficient wherever it runs. This work is early and exploratory; it is not a product and makes no performance claims.
Measured, not modeled
Bench results
November 2026
Measured on development hardware, not modeled.
A complete open-source language model runs with every feed-forward layer on our low-power accelerators, within 1% of full-precision accuracy on a standard benchmark.
Larger models lose even less: 1.5B to 7B models lose 0.1–0.4% at our precision (software simulation; silicon runs next).
Serving many requests together changes nothing: answers are bit-for-bit identical, however the work is batched.
Under ~1 W per accelerator chip on language-model layers (an upper bound: below our meter’s floor).
What we’re working on now: larger models on silicon, and the memory that serves many conversations at once.
Want the details? We share architecture and test data with researchers, partners and investors under NDA. Email us →
Where we are
A project in motion.
Figures on this page are design targets and will be updated as modeling and testing progress. The Bench results section is measured.
- NOW · IN PROGRESS
Architecture & modeling
Gen‑3 panel design iteration underway, with more iterations expected. Now: bill of materials, panel simulation and bench tests of key parts.
- NEXT
Ground validation
Build and measure tiles and a tile group on the ground: power, heat, cold survival, data links and useful output.
- LATER
Vela-Scout 83 rideshare
A proposed SpaceX Falcon 9 rideshare demonstrator and our current objective. Not yet committed.
Latest updates
- Published Paper v1: The Panel Is the Computer (PDF).
- New headline benchmark: tokens per solar watt, with H100 equivalents, November bench results, and a new page: how we get there.
- Published our accepted-work KPIs, panel design targets and power-gap framing.
- First bench results published, measured on development hardware.
- Panel design reached its third generation (Gen‑3); further iterations expected.
- Project website and overview video launched.
Get involved
Looking for builders.
Project Vela is assembling a small team of builders in power electronics, thermal, semiconductors, and spacecraft systems.
If that’s you, or you know someone, get in touch: info@project-vela.com · @ProjectVela
Solo founder, looking for a business partner / operator. About · Careers · FAQ · Press
Our approach
Start small.
Measure the whole system.
Progress means replacing assumptions with evidence, one experiment at a time.
- 01 / DEFINE
Make the question testable.
Choose a concrete workload and specify what success—or failure—would look like.
- 02 / MEASURE
Build the smallest useful test.
Include memory, power, data movement, and recovery costs. Compare against a credible baseline.
- 03 / REFINE
Let evidence set the direction.
Keep what demonstrates value. Revise or stop approaches that do not survive testing.
Current stage Preliminary engineering and research design. Panel design is at its third generation (Gen‑3); bench validation is the next milestone.
Is this an operating satellite service?
No. Project Vela is an early-stage research and engineering initiative. It has not demonstrated an operating orbital compute service. Vela-Scout 83 is proposed as a rideshare demonstrator, but no flight is committed.
Which chips does Vela use?
Vela uses mature, high-volume chip processes that avoid the EUV bottleneck and its cost premium, and it targets inference rather than training. Specific parts are not published. Selection will be driven by measured useful output per dollar and per kilogram for the complete system.