ScitiX is not an easy sell on GPU price: verified H100 hours run $1.50–$12.41 across official providers, and ScitiX publishes no comparable rate
Asked (summary)
Thinking of joining scitix.ai/gpu-cloud as an enterprise AE — is this competitive pricing and easy to sell?
Evidence: 100 official price rows from AWS, CoreWeave, Google, Lambda, Nebius and Crusoe (deduplicated to 42 distinct offers, normalized to USD per GPU-hour by dividing cluster prices by GPU count), plus 44 ScitiX offer extractions and 30 company-signal extractions. No verified ScitiX per-GPU-hour price was found on the linked page; its public claims trace largely to one syndicated press release. Research date: 10 September 2026.
Benchmark: what one GPU-hour costs at official providers
Hover a dot for the exact SKU, GPU count and region. One AWS SageMaker B200 row ($113.67 per instance-hour, GPU count unstated) is excluded from the chart as non-normalizable; it appears flagged in the table.
What the numbers say
H100 on-demand spans $1.50 (Nebius no-commitment) to $12.41 per GPU-hour (Google H100 Mega) — an 8× spread inside "official" pricing alone, so any ScitiX quote must state SKU, term and region before it can be called cheap. nebius.com
CoreWeave spot cuts its own on-demand by roughly 60%: H100 $2.44 vs $6.16, H200 $2.58 vs $6.31, B200 $4.26 vs $8.60 per GPU-hour — spot and on-demand must never sit in the same comparison column. coreweave.com
ScitiX's own site claims "30–40% lower cost" but publishes no per-GPU-hour rate; a Singapore accelerator lists a 12-months-for-the-price-of-11 contract, and one outlet cites per-token inference prices instead. Opaque pricing is itself a sales-friction signal. scitix.ai
ScitiX's traction claims — 1 trillion tokens daily, ~1s time-to-first-token, >90% cache hit, 99.9% uptime, RadixArk/SGLang as customer — appear across ~28 domains but trace to one press release, which itself says the metrics "do not constitute service-level guarantees". aijourn.com
Sellability scorecard
Weak
Raw GPU price competitiveness
No verified public per-GPU-hour rate. Nebius and Lambda already sell H100 at $1.50–$3.99; competing on sticker price against them is a losing pitch without a proven price book.
Weak
Independent proof
One named user (RadixArk); traction figures are self-reported, syndicated, and explicitly disclaimed as non-guarantees. No audited references in the data.
Credible
Differentiated enterprise inference
Owned B200/H200/H100 fleet, dedicated tenancy, zero-retention, model routing and fallback, context reuse, audit logs, 99.9% uptime SLA — a managed-platform story, not commodity rental.
Unproven
Company maturity
A company database lists ScitiX as unfunded, founded 2025, Singapore-based, while press releases date it San Francisco; a job post signals it is only now hiring its first US technical hire.
Verdict and interview diligence checklist
High-risk, high-upside. Join only if management can show at least 2–3 referenceable enterprise customers, deliverable GPU inventory, and a quota built on current — not planned — capacity. Verify before signing:
Actual price book and discount floors; gross margin after power, network and egress
Sellable capacity by GPU model and region today, and whether you sell raw GPU hours or the managed inference platform
Signed customer references beyond RadixArk; renewal and expansion rates
Median ACV, sales-cycle length, pipeline coverage, quota and attainment by rep
SLA credits, 24/7 support, SOC 2 / ISO / HIPAA and data-residency readiness, private networking
All 42 normalized benchmark offers
Provider
GPU
SKU as listed
$ / GPU-hr
GPUs
Type
Region
Source
Method: 100 official-provider price rows collected 2026-09-10 from AWS, CoreWeave, Google/Google Cloud, Lambda, Nebius and Crusoe pricing pages, deduplicated to 42 distinct offers; cluster and instance prices divided by stated GPU count to get USD per GPU-hour. Spot, on-demand, reserved/capacity-block and promotional offers are kept separate; rows with unclear unit or GPU count (GKE management fees, Colab compute-unit conversions, a SageMaker instance price) are flagged, not treated as full GPU cost. ScitiX claims come from 44 offer and 30 company extractions dominated by one syndicated press release; no verified ScitiX GPU-hour price exists in the data. Prices are indicative and exclude storage, networking and egress.