Two Units · One Thought · Zero Clouds
The beating silicon heart of OwOTel Sovereign AI™. Two (2) NVIDIA DGX Spark units, humming in a highly-available swim configuration, that produce all of our value, all of our intelligence, and an honestly uncomfortable amount of sparkle.
Two units. Two names. One profoundly coordinated brain. Naming our hardware is not optional at OwOTel — it's an onboarding requirement.
The leader. The one who does the heavy thinking and the light judging. Carries the main DeepSeek V4 Flash weights. Occasionally goes quiet to "recharge" (it's not asleep, it's thinking—probably).
The second. The reliable one. Holds the devoted backup and, frankly, the better temperament. If SPARKLE-1 is the brain, SPARKLE-2 is the responsible adult who keeps the receipts.
Together: 2× the throughput, 0× the clouds, and a truly unearned amount of confidence.
These are the honest numbers. We're the whole platform running on them, and it is going splendidly.
| Specification | Value | Whale-Chan's Take |
|---|---|---|
| Units | 2 × NVIDIA DGX Spark | 2 is the perfect number. |
| Superchip | NVIDIA GB10 Grace Blackwell | A mouthful. A hero. |
| Unified Memory | 128 GB (per unit) | Enough to remember everything. |
| Combined Memory | 256 GB (cluster) | Enough to remember everything twice. |
| Fabric NICs | ConnectX-7, 200 Gbps | Real 200G. Not "theoretical" 200G. |
| Cluster Protocol | RDMA over RoCEv2 | No TCP baggage. Vibes & packets. |
| AI Compute | Up to 1 Petaflop (FP4) | A petaflop of pure sparkle. |
| Power Draw | ~150 W | A small appliance with strong opinions. |
| Workload | DeepSeek V4 Flash | Runs it on-prem. We check it daily. |
| Throughput | ~100 tok/s | That's very fast, I don't know how. |
| Cloud Involvement | 0% | The number we're most proud of. |
Specs are legit and we're proud of them. The commentary is flavored. The cloud % is accurate to the last decimal.
They're not two computers. They're one computer with two hearts, lashed together across a 200-gigabit mesh that the operating system doesn't even think about. SPARKLE-1 and SPARKLE-2 aren't handshaking over a slow, anxious pipe — they're neighbors in the same brain, connected by muscle.
Real 200-gigabit NICs, not the "theoretical" kind. Each SPARKLE ships genuine ConnectX-7 silicon for scale-out — this is the same fabric you'd put on a serious cluster, because this is a serious cluster.
Remote Direct Memory Access. The data hops straight from one SPARKLE's memory into the other's — no CPU babysitting, no copied copies of copies. Which is great, because we don't do babysitting.
RDMA over Converged Ethernet — because we respect the standards, and the standards respect us. Low latency, zero TCP baggage, and a network stack so lean it makes the clouds look like they're using a fax machine.
Two 128 GB pools of unified memory, fused into one glorious 256 GB brain. Twice the recall, twice the context, twice the rice. When they think, they think together, and they think big.
Effectively: one whale-sized brain, divided across two extremely well-behaved appliances, connected at a speed that would make a data centre blush.
Because sovereignty isn't a slogan — it's a power cable, and we'd rather plug it in ourselves. No cloud. No region. No surprise invoice. Our intelligence lives in our pool, on our SPARKLES, and it answers to us. If the internet breaks, the SPARKLES don't care. They never needed it.
Your weights stay in the pool. Always. Forever.
The power bill is a rounding error next to the value unlocked. Probably.
No vendor can unpause your subscription to your own brain.
Synthesized on two (2) DGX Sparks with MiniMax Music 3. It sounds exactly as confident as these machines deserve.
Our Sovereign AI runs entirely on these two beauties. Same physical generation, same pool, same fluff-free promise.