V-2
More tokens, less capexMore tokens, less opexSame eight GPUs. More tokens every hour, and every watt.
Less capex
One hour. All eight GPUs. Full load.
Nova 8cooled by Nexalus, tuned clocksMeasured
194,342,791
53,984 tokens/s
Tier-one air-cooledreference server, publishedReported
163,897,200
45,527 tokens/s
56k48k40k
air-cooled reference 45,527
+30.4 M extra tokens
hottest GPU 56 °C
80 °C, output held
+19% tuned · +14% out of the box Measured vs reported
Buy 1,000 Nova 8s, get the output of 1,186 air-cooled servers
Estimate
V-5 →
One CPU, single-slot GPUs Design
1 CPU for eight full-power GPUs (the tier-one reference uses 2). Each GPU takes one slot, not two.
Air2 slots
Nexalus1 slot
Investor read Nearly a fifth more output from the same silicon lets us price on the value we create.
Less opex
Cooler cards ran faster.
+22%tokens per watt
≈ 31 M vs ≈ 25 M tokens per kWhEstimate
≈ 31 M vs ≈ 25 M tokens per kWhEstimate
In one box, the coolest card ran 9.6% faster than the hottest (six cards, stock clocks). Card-to-card variation is not yet separated from temperature; a same-card sweep is planned. Measured
68 °C
6,534 tokens/s per card
+9.6% vs 87 °C
Cooling power at full speed Estimate
Nova 8 pumps + fans~800 W
Reference fans~2,100 W
Fan and pump ratings at full speed (not metered): Nova 8 ~800 W budget; reference fans up to ~2,100 W nameplate. Operating draw is lower for both.
Investor read Grid power caps the AI build-out. More tokens per watt sells wherever the grid is tight.
V-1, measured: Nova 8 one-hour run, 5 Oct 2026, 175 back-to-back runs on a tuned development build (8 × RTX PRO 6000 Blackwell at 600 W, Llama 3.3 70B, NVFP4, TensorRT-LLM): average 53,984 tokens/s, peak 54,968 on run 1; the hottest GPU rose from 56 to 80 °C while output fell about 2%. Tuned = core +600 MHz, memory +1,400 MHz, clocks locked at 2,550 MHz. Load settings differ from the published reference (ours 256/256 tokens, 512 requests per GPU; reference 2 replicas per GPU, 8,192 users). A same-settings run is scripted. Out of the box (stock clocks): 51,705 tokens/s, +14% (measured). Reported: 45,527 tokens/s published for the tier-one air-cooled 8-GPU reference server on the same GPUs; the comparison is against the published result, not a same-silicon A/B. The race replays the measured run, 60 minutes in 12 seconds; the page rests on the finished hour. 1,186 = 1,000 × 53,984 ÷ 45,527 (estimate). One CPU: the built Nova 8 runs eight full-power GPUs from one CPU; the tier-one reference uses two. Single-slot: each liquid-cooled GPU takes one slot where an air-cooled card takes two (design). V-2, estimate: +22% tokens per watt = 54,968 tokens/s (run 1) ÷ ~6.4 kW estimated draw (two of four PSUs metered) vs 7.03 tokens/s per W published for the reference, so 8.6 vs 7.0 tokens/s per W ≈ 31 M vs 25 M tokens per kWh; +20% on the one-hour average. GPU-only 11.5–11.8 tokens/s per W measured. Measured dots: six cards in one Nova 8 development unit, one run, stock clocks (Aug 2026): the 68 °C card made 6,534 tokens/s and the 87 °C card 5,961 (+9.6%). Six cards cannot separate card-to-card variation from temperature; a same-card temperature sweep is planned. The slider follows a straight trend through the end cards (illustrative); beyond 68–87 °C it is extrapolated. Cooling: fan and pump ratings at full speed, not metered: ~800 W budget for Nova 8 pumps and fans vs up to ~2,100 W fan nameplate in the reference. Operating draw is lower for both.
V-1 · V-2 · More tokens