How TensorArt, SeaArt and RunningHub obtain GPU capacity#

Status: Decision log Research date: 2026-10-03 Scope: Internal competitive research; procurement and operating models, not a complete hardware inventory. Decision boundary: This report does not reopen the founder's permanent ownership goal, buying decisions or three capacity lanes.

2026-10-03 — Findings and competitive implications#

Answer#

These platforms do not all follow the same model. The important distinction is who operates the infrastructure versus who legally owns the GPUs. Self-hosting and renting can coexist.

Competitive conclusion: Owning GPUs gives us control over our base-capacity economics. It does not establish a unique infrastructure advantage against TensorArt's disclosed self-operation, and RunningHub is not merely an application buying retail serverless inference. Against SeaArt, avoiding a cloud supplier's margin is a plausible advantage—not a demonstrated cost or latency lead.

What “own infrastructure” does and does not mean#

Four different arrangements matter:

• Purchased GPUs, self-operated: hardware capital is paid upfront; the operator controls workers, storage and scheduling.
• Dedicated leased servers, self-operated: someone else owns the hardware, but persistent workers and model caches can still be operated by the lessee. Already-committed idle time may be usable without another per-job hardware charge; actual contract terms govern.
• Cloud GPU instances: infrastructure is rented, but applications can keep instances warm, retain caches and optimize scheduling. Spot instances introduce interruption/availability tradeoffs, not an obligation to cold-start every request.
• External model services: the platform integrates somebody else's inference service. Owning local GPUs does not eliminate those upstream costs or dependencies.

These are analytical categories. A data-center claim does not establish ownership of the building, land or every installed GPU. Likewise, a “cloud ComfyUI” product description tells us where users execute work, not how its provider procures hardware.

TensorArt: direct evidence of in-house operation#

In the interview published by Z Potentials on 2025-04-15, founder Zhenyu Shen explains its inference-price strategy:

“we built our own data centers and manage a large number of GPUs in-house”

He also attributes lower compute costs to this infrastructure and inference engineering, claiming a fivefold advantage against Civitai. That comparison is self-reported, with no matched workload, accounting boundary or independent verification; it is not a usable estimate of our advantage. The disclosure nevertheless directly resolves the operating-model question: TensorArt has built and operated an inference infrastructure layer, rather than relying solely on an opaque third-party inference API. Original founder interview, section 04.

The picture is not “everything is in-house.” TensorArt's 2026-03-31 incident notice attributes Nano Banana disruption to GCP while saying other features remain operational. Its 2026-04-13 Canvas announcement lists closed-source models and a separate Energy resource system. This establishes an external dependency for that feature and a distinct commercial-model offering; it does not locate the entire open-model fleet on GCP. API/service integration is the reasonable inference, not proof of a particular upstream contract. TensorArt's own notices.

Not publicly established here: fleet count, serving GPU SKUs, procurement vendors, leases versus purchases, financing, locations, and October 2026 allocation. A GPU mentioned in a model-training card is not evidence of the serving fleet.

SeaArt: rented GPU inference, with substantial cloud-side optimization#

Google Cloud — explicit GPU evidence. Its customer case identifies NVIDIA L4 GPUs attached to Spot VMs, managed through GKE, for SeaArt image generation. It describes adding/removing resources and reports a 30% infrastructure-cost reduction. This is a named workload, accelerator and rental mechanism—not merely cloud storage or web hosting. The page has no clear publication date and describes early growth following its 2023 launch; its old usage figures must not be treated as today's fleet scale. The animation/video expansion described there is a plan, not proof that current video jobs use L4s. Google Cloud customer case.

AWS — another documented provider, not an exclusive-provider claim. The official 2024 case lists EC2 and SageMaker alongside other services and says generative-AI drawing runs on AWS. Bedrock is described for prompt optimization and customer support; that alone would not prove the image-generation GPU supplier. The case supports additional cloud deployment without specifying image-serving GPU SKUs or a present workload split. AWS customer case.

Alibaba Cloud — directly relevant to our cache/latency thesis. A 2024-10-14 account of an Alibaba product leader's presentation describes SeaArt's containerized inference and Fluid hot-model caching, reporting 8–10-second average image generation. This establishes that remote model storage and cold loading are actively engineered around, rather than unavoidable characteristics of rental. Alibaba presentation account.

A newer 2025-07-22 post from Alibaba's infrastructure account describes a DeepGPU deployment, reporting 3.95-second inference and over 50% end-to-end improvement, including model/LoRA switching improvements. These are vendor-reported results with unspecified workloads, not an apples-to-apples comparison to our Krea, WAN or MiniMax runs. They are still strong evidence against assuming SeaArt has an unoptimized generic cloud path. Alibaba customer story.

Resolved: SeaArt has obtained real image-generation GPU capacity from public cloud. Not resolved by public disclosure: whether it owns any other GPUs, which provider currently serves each region/model, or current cloud discounts and commitments. The dated cases cannot establish simultaneous October 2026 traffic shares. “Entirely cloud-only today” would overstate the evidence.

RunningHub: Haima's infrastructure business, with wholesale rental#

Product-to-infrastructure link. RunningHub's own about page identifies 安徽海马云智能科技发展有限公司 as its owner. Haima's website lists RunningHub among its native AI applications and describes a self-developed GPUaaS platform with mixed real-time rendering and inference scheduling, GPU clusters and its storage/control stack. This is direct evidence of an infrastructure operator behind the application. It does not establish legal ownership of all machines. RunningHub about, Haima Cloud.

Named bulk supplier. A 2025-07-17 announcement attributed to Hainan Holdings/Hainan Huatie says Haima will rent large-scale GPU resources from Hainan Huatie, including GPU servers, management servers, storage and support. It gives an aggregate RMB 1 billion rental/service scale. A supplier investor-relations record published 2025-08-13, page 2/Q1, also acknowledges the agreement. This is group-level contracted capacity, not an annual GPU bill, delivered inventory, or RunningHub-only expenditure; the term, GPU SKUs and application allocation are not supplied. Supplier-authored announcement reproduced by Sina, company investor-relations record.

UCloud — verified history, important current correction. UCloud's 2026-06-16 regulatory response confirms server-rental and IDC relationships with Haima. Crucially, its contract table marks the older server-rental arrangements completed; its explanation says some contracts were terminated early and current cooperation is limited to a small amount of IDC business. Therefore “RunningHub currently rents its GPUs from UCloud” is not a supported conclusion. The filing also describes the relevant historical architecture as ARM-oriented: it cannot identify RunningHub's AI accelerator fleet. UCloud regulatory response, question 3.

Broader procurement, lower evidence tier. March 2026 reporting on Haima's IPO draft identifies server manufacturers, compute lessors and IDC operators, and quotes Haima treating rented servers as controlled right-of-use assets. This supports a hybrid procurement ecosystem; it is not an audited owned-versus-leased GPU count. The linked HKEX application PDFs returned unavailable/404 during this research, so this report does not pretend to have inspected the full draft. Purchasing from a server manufacturer, supplying servers to customers, operating an IDC, and owning GPUs retained for RunningHub are different claims. Securities Market Weekly reporting.

Resolved: RunningHub sits on a vertically integrated, self-operated GPU platform whose parent uses wholesale rental. Do not claim: it owns 100% of its GPUs, rents 100%, buys every job from a retail serverless provider, or runs on a verified count/model of GPUs. A “mixed owned and leased fleet” is plausible, but the owned share is not conclusively established by the sources inspected here.

What this changes for ComfyICU#

The founder's ownership goal remains intact. These findings change the competitive premise, not the decision to own.

The strongest remaining opportunity is execution of a narrowly targeted service advantage:

1. Make the owned fast lane reliably fast for its resident model families. Control over cache, workers and admission is valuable. Compare the same workflow, resolution, steps, precision and output count—including queueing, failures and artifact delivery—before claiming a competitor lead.
2. Use idle owned time to create demand while protecting paid work. Preserve the existing short/preemptible gap-fill rules, the free-node reserve and no-Modal-spill boundary. The advantage is the result—retained users and dependable paid latency—not the mere existence of free credits.
3. Compete on the right cost boundary. Compare fully loaded cost per successful equivalent job at observed utilization, including power/cooling, capital consumption, maintenance, failures, storage/network and overflow. Separately track incremental gap-fill cash cost and its opportunity cost. Competitors' Spot prices, negotiated leases and throughput optimizations rule out a defensible numerical cost advantage from public information alone.
4. Keep the rented bridge for work outside the local lane. Long/bursty jobs and larger-memory work remain separate. External proprietary-model services also do not become electricity-only workloads because we own GPUs.

Haima's mixed scheduling makes cross-workload utilization a credible competitor capability; whether it actually transfers spare gaming GPUs to RunningHub is not disclosed. TensorArt's interview describes creator/model-network advantages alongside infrastructure. Our smaller fleet must be assessed against optimized operators, not an assumed unoptimized serverless reseller.

The owned-GPU strategy records three purchased RTX 5090 nodes and pending production integration as of its September evidence. This investigation did not query the live scheduler or change that deployment state. Hardware capability must become a measured customer result before it is advertised as a market lead.

Evidence quality, reproducibility and remaining limits#

Research combined English/Chinese searches, original interviews, official platform pages, cloud-provider customer cases, supplier disclosures and read-only headless-Chrome retrieval. Search snippets and AI-generated summaries were discovery aids, not load-bearing evidence. All cited sources were read on 2026-10-03; supplier case-study metrics are reported claims, not independently benchmarked results.

The important sources were archived with the local research CLI. The source manifest records public URLs, dates, retrieval mode, run IDs and immutable raw S3 archive locations. Raw archives remain outside the report commit; derived page Markdown is also retained locally under raw-archive/website/fetch/. Browser-rendered captures were used where HTTP returned only an application shell. The supplier PDF was downloaded and its page-2 agreement checked directly; its checksum is recorded in the manifest.

Confidence is high for positively disclosed operating/procurement mechanisms, not for an exhaustive current inventory. Specifically:

• TensorArt's self-operation disclosure is from April 2025; no current purchase ledger was disclosed.
• SeaArt's provider cases prove real deployment at their stated times, not exclusivity or current percentages.
• Haima's supplier agreement proves the wholesale-rental route, not delivery of the entire announced capacity to RunningHub.
• UCloud's newer disclosure takes precedence over earlier reporting when describing that supplier relationship today.
• No inspected source provides all three companies' utilization, lease prices, fleet title or comparable job-level costs. “Unknown” here is a bounded disclosure limit, not a substitute for the positive findings above.

The strategic answer is conclusive at the useful level: the market contains self-operated infrastructure, public-cloud rental, and self-operated platforms backed by wholesale leases. We cannot base our competitive advantage on all three rivals being simple GPU renters with unavoidable cold starts.