ScitiX Bets Owned NVIDIA Clusters Win Multi-Model Inference

ScitiX launched its full production inference platform on August 16, 2026, running entirely on company-owned and operated NVIDIA B200, H200 and H100 clusters and already processing more than 1 trillion tokens daily with average time-to-first-token near one second, cache hit rates above 90 percent and 99.9 percent uptime.

The Singapore-based company is placing a clear wager: enterprises moving multi-model and agentic workloads into live traffic will pay for a neutral execution layer that owns the GPUs, routes intelligently and enforces compliance, rather than shopping solely for the lowest token price among crowded API hosts.

One Trillion Tokens a Day on Bare Metal

The platform, called ScitiX Model Inference, targets organizations that already run several open-source, fine-tuned or third-party models at once. It presents a single familiar API surface and sits as a routing and operations layer above the hardware.

Current production numbers published with the launch include:

• Over 1 trillion tokens processed every day
• Approximately one second average time-to-first-token
• Cache hit rate exceeding 90 percent
• 99.9 percent uptime

Those figures come from live customer traffic, not lab demos. The company states the entire stack runs on ScitiX-owned infrastructure rather than rented cloud slices.

What Ships Inside the Execution Layer

ScitiX lists six core capabilities that move the product beyond simple model hosting:

• Intelligent model routing and fallback that picks the best model for latency, cost or quality targets and fails over automatically
• Session-aware context reuse that keeps long conversational state and caches intermediate results to cut redundant compute
• Fault-tolerant execution that absorbs retries, timeouts and partial failures without breaking the whole workflow
• Private deployment environments with dedicated tenancy for data residency or security rules
• Zero-retention policies so prompts and outputs do not persist past the transaction
• Full-stack observability with infrastructure telemetry, audit logs and dashboards that show exactly where latency or cost originates

The company frames these as the difference between experimental calls and production systems that must survive real traffic variance.

Routing That Treats Models as Interchangeable

Enterprises no longer run one model forever. Agentic pipelines can fire dozens of internal inference calls per user action. ScitiX’s bet is that the layer choosing which model answers, when to fall back, and how to reuse context becomes more valuable than any single set of weights.

The platform keeps the same endpoint and auth while the model name changes. That design lets teams swap open-weight families or fine-tunes without rewriting client code. Cache-aware session reuse is presented as the main lever for cost control once token prices themselves have already fallen.

Live per-token pricing across 21 models is published openly. Examples include DeepSeek-V4-Flash at $0.14 input and $0.28 output per million tokens (cache hit $0.028), Qwen3.6-27B at $0.060 / $0.40, and larger options such as DeepSeek-V4-Pro at $1.74 / $3.48. Batch calls run at half price on eligible models. Cache-hit pricing applies to cached input tokens.

The table is a sample of the public catalog; ScitiX also lists embeddings, vision and speech models.

The SGLang Team Already Runs Heavy Workloads Here

Proof points matter more than feature lists. RadixArk, the commercial company behind the popular open-source SGLang high-performance serving framework, runs its heaviest production scenarios on ScitiX.

As inference gets more complex, the underlying infrastructure becomes the differentiator. ScitiX delivers the responsiveness and reliability we depend on.

RadixArk said that in the launch materials. SGLang itself grew out of UC Berkeley work and spun into a funded commercial entity earlier in 2026 with a large seed; its choice of host gives the unfunded infrastructure provider a visible reference customer in the exact niche of low-latency, high-throughput serving.

Separately, technical discussion around ScitiX has highlighted FlashBoot, a weight-loading approach built with SGLang that packs parameters into contiguous CUDA arenas and pipelines clones across the fabric. Claims circulating for a 1.6-trillion-parameter replica reaching serving-ready state in roughly 0.32 seconds illustrate why co-design between the framework and bare-metal fabric can matter for autoscaling and fault recovery far more than raw peak FLOPS.

How Cache Hits and Zero Retention Change the Bill

Token sticker prices keep dropping across the industry. Total spend often does not, because every user turn can fan out into many internal calls. ScitiX’s published cache hit rate above 90 percent is the operational claim that turns that cascade into something closer to a controllable line item.

Zero-retention and private tenancy options address the procurement reality that many regulated buyers now treat data terms as non-negotiable. When a frontier model provider changes terms with a new release, third-party API customers can discover the shift after the fact. Owned infrastructure plus explicit no-persist policies is ScitiX’s answer to that sovereignty requirement.

Internal tooling called SiEval goes beyond leaderboard scores. It examines the full execution chain for reproducibility and decision support (release, rollback, checkpoint). In internal tests ScitiX reported up to 10.5× acceleration on evaluation-heavy pipelines and 7.22× end-to-end speedups on large leaderboard workflows, with the biggest gains on LLM-as-judge, sandboxed code and long-context jobs.

Where the Unfunded Stack Meets Billion-Dollar Rivals

The inference market is crowded with well-funded names. Baseten has raised multi-billion-dollar rounds at high valuations. Fireworks, Together, Groq and others compete hard on token price and catalog breadth. Deployment-focused platforms sell control of the replica lifecycle.

ScitiX appears in databases as an unfunded company founded around 2025 and headquartered in Singapore, yet it claims ownership of current-generation Blackwell and Hopper clusters and already carries production volume that many funded peers would advertise. That combination is the core of the bet: capital efficiency through ownership and specialization rather than repeated large raises for shared capacity.

Hardware context helps. NVIDIA positions its DGX B200 systems as delivering 15X the inference performance of prior systems under certain projected comparisons. Owning and tuning that class of gear, instead of renting slices, is how ScitiX argues it can guarantee the latency, cache and observability numbers it publishes.

Enterprises already making large internal AI bets, including Microsoft’s large internal AI engineering push, face the same multi-model operations problem at greater scale. Reliability failures such as recent global ChatGPT and API outages also keep the value of dedicated, observable infrastructure visible.

Control Levers Enterprises Buy

ScitiX’s public statement is direct: “We are not building another model. We are building the operational layer that makes multi-model production viable. Enterprises should not have to become GPU operators to deploy AI. They need flexibility, control, and a platform that handles the rest.”

The product is available now for enterprise customers. Deployment options, full pricing and the supported model list sit on the company site. Performance metrics and customer statements are presented as illustrative; actual results depend on workload, model architecture, configuration and network conditions.

The wager is that as agentic systems multiply internal calls and as compliance requirements harden, the buyer who cares about per-request cost, latency SLAs, auditability and model agility will prefer a neutral, owned-stack operator over both pure model vendors and pure token discounters. Early volume and a demanding reference customer give the claim something concrete to stand on.

Frequently Asked Questions

What hardware does the ScitiX inference platform run on?

The entire platform runs on ScitiX-owned and operated NVIDIA B200, H200 and H100 clusters rather than multi-tenant public-cloud slices, giving the company direct control over scheduling, cooling and interconnect tuning.

What production metrics did ScitiX publish at launch?

Alongside the August 16, 2026 announcement the company cited more than 1 trillion tokens processed daily, average time-to-first-token of approximately one second, cache hit rates exceeding 90 percent, and 99.9 percent uptime drawn from live traffic.

Who is already using ScitiX for heavy inference?

RadixArk, the commercial team behind the SGLang serving framework, runs its most demanding production scenarios on the platform and publicly praised the responsiveness and reliability of the underlying infrastructure.

Does ScitiX publish model pricing?

Yes. The public catalog lists per-million-token input, output and cache-hit rates for roughly two dozen models including DeepSeek, Qwen, MiniMax, Gemma and gpt-oss variants, with batch discounts at 50 percent on eligible models.

Is ScitiX a funded startup?

Company databases list ScitiX as an unfunded firm founded in 2025 and based in Singapore, even while it operates current-generation NVIDIA clusters and carries multi-model production volume.