Last Updated on July 30, 2026 by metanetdev
Tenstorrent AI Servers | Wormhole, Galaxy & Dedicated AI Hosting
For most of the last decade, purchasing AI infrastructure meant choosing a conventional GPU server and building around a familiar CUDA software stack. That remains the right answer for many workloads. But AI infrastructure is becoming too important—and too expensive—for businesses to assume that one hardware architecture must be the answer to every inference problem.
Tenstorrent AI servers give companies another credible path. Metanet offers available Wormhole servers, Galaxy systems, and other Tenstorrent configurations as dedicated AI hosting products for organizations that want scalable inference infrastructure, an open development environment, and an alternative to treating public-cloud GPUs as the only way to run AI.
Metanet web hosting division offers Tenstorrent systems from available stock alongside NVIDIA H100, H200, B200, and B300 infrastructure and AI colocation in NYC and New Jersey. That matters because customers can choose the best platform for their actual model, software, capacity, and budget—not be forced into a one-size-fits-all GPU lease.
Tenstorrent systems available through Metanet
Metanet can position the product line according to the customer’s stage and compute requirement. Exact availability should be confirmed during quoting, but the key deployment choices are clear:
| Tenstorrent platform | Best fit | Hosting message |
|---|---|---|
| Wormhole servers | Compatible AI inference, proof-of-concept deployments, and dedicated alternative-accelerator capacity | Available dedicated AI hosting capacity for customers who want to validate and deploy now |
| Galaxy systems | Larger multi-accelerator deployments and customers planning a serious AI compute environment | A high-density Tenstorrent server platform for scalable AI infrastructure, networking, and growth |
| Wormhole PCIe systems | Smaller deployments and configuration-specific projects | Flexible entry point for supported workloads and custom server designs |
| Blackhole systems and cards | Customers evaluating newer Tenstorrent hardware and greater local memory or connectivity | A forward-looking platform for customers planning their next AI server generation |
This range lets Metanet serve the customer who needs a single dedicated AI server as well as the company evaluating a multi-accelerator platform. The qualification process starts with the model, the expected traffic, and the desired operating environment—not with a generic hardware quote.
What is Tenstorrent Wormhole?
Wormhole is Tenstorrent’s AI accelerator architecture. Rather than being a general-purpose graphics processor repurposed for AI, it is designed around AI compute and efficient data movement. Tenstorrent’s approach combines its Tensix cores, on-chip memory, local high-speed memory, and an architecture designed to scale processors together through Ethernet-based connections.
At the individual-card level, Tenstorrent documents Wormhole PCIe products with 80 Tensix cores, 120 MB of SRAM, and 12 GB of GDDR6 memory. Tenstorrent Wormhole PCIe documentation The cards can be used in suitable systems, while larger Wormhole server platforms let customers use many accelerators as one AI compute environment.
The important point is not the component count. The important point is that Wormhole is a serious AI infrastructure alternative for workloads that are compatible with the Tenstorrent platform. It gives AI developers another way to think about performance, cost, system design, and supplier concentration.
Dedicated Tenstorrent AI hosting versus public cloud GPU instances
Public cloud remains useful for testing models, handling seasonal spikes, and using managed services. But an AI company with a sustained inference workload needs to ask a different question: how much dedicated compute is required every month, and how much does that compute cost once networking, storage, data transfer, and reliability are included?
Dedicated Tenstorrent AI hosting can give the business:
- A known server configuration reserved for its exclusive use.
- Predictable monthly capacity rather than an open-ended hourly bill.
- Direct control over the operating system, container environment, network, and storage design.
- A platform for measuring a real workload instead of relying on generic benchmark claims.
- An alternative accelerator path for supported AI models.
- A practical way to test non-NVIDIA infrastructure without buying an entire private cluster first.
This is especially relevant for AI inference. Once an application is serving a stable model repeatedly—whether it is an internal assistant, document-processing workflow, vision platform, or customer API—the workload can justify dedicated infrastructure. A fixed server lets engineering teams tune their model, batch sizes, quantization, storage path, and network behavior against a known environment.
Where Wormhole can be a strong fit
Wormhole should be evaluated based on the application, not hype. Its strongest use cases are generally inference-oriented workloads and development programs where the team is prepared to use Tenstorrent’s software ecosystem. The platform may be particularly interesting for organizations that want to explore efficient AI compute without assuming the only path is an expensive conventional GPU deployment.
Potential uses include:
- Large language model inference for supported model families.
- AI chat and enterprise-assistant applications.
- Retrieval-augmented generation, where the infrastructure combines inference with vector search and document stores.
- Computer vision, image classification, and video-analysis workflows where models have been qualified for the platform.
- Model experimentation and accelerator evaluation.
- AI service providers looking to offer differentiated infrastructure choices to customers.
- Private AI deployments where the customer wants dedicated hardware and a controlled software environment.
The right way to sell this is simple: bring the workload, test it, and measure it. Throughput, latency, model support, batching behavior, precision, memory requirements, and engineering effort should all be evaluated on the actual application. That is much more valuable than repeating a broad claim that one accelerator is universally superior to another.
Open software is a meaningful part of the proposition
Tenstorrent’s hardware story is closely tied to its software approach. The company supports open-source developer environments including TT-Forge for higher-level development and TT-Metalium for lower-level programming. Tenstorrent Wormhole product overview
For some teams, that openness is strategic. It can make the accelerator platform easier to inspect, adapt, and optimize rather than treating every critical software layer as a black box. It may also appeal to engineering organizations that want a closer relationship with their inference stack and are willing to invest in platform expertise.
That is not a claim that the Tenstorrent path is effortless. CUDA has a very large installed base, familiar tools, and broad third-party support. A company whose application depends on a CUDA-only library or an unported model should not pretend the transition is automatic. But for a team that can validate its workflow, Tenstorrent’s open environment can be an important reason to look beyond a single accelerator vendor.
Wormhole servers scale differently from ordinary GPU hosting
AI buyers often focus only on individual-card memory or benchmark output. In a real server deployment, the system architecture matters just as much. Model parallelism, data movement, interconnect behavior, host CPU, storage, network interfaces, and software scheduling all affect results.
Wormhole’s scale-out orientation is one of its distinguishing characteristics. Tenstorrent describes its processors as capable of networking into multi-chip configurations, with the wider platform designed around data movement and mesh-style scaling. This makes it relevant to customers who are thinking about an AI service as a system, not merely a single GPU rental.
Metanet can help scope the entire hosting design: server count, storage, public or private network exposure, management access, IP address plan, DDoS policy, and growth path. An AI hosting deployment should be built so a customer can begin with a practical initial footprint and expand after the workload proves itself.
Tenstorrent Wormhole and NVIDIA: choose by workload
The most credible positioning is not “Tenstorrent replaces NVIDIA for everyone.” It is “Metanet offers customers a serious choice.” NVIDIA H100, H200, B200, and B300 systems are strong options for organizations that need CUDA compatibility, established frameworks, high-memory configurations, and mature multi-GPU infrastructure.
Wormhole is a differentiated option for organizations that want to evaluate an alternative AI accelerator, prioritize compatible inference performance and cost, or build on a more open software foundation. A customer may choose one or the other—or use both for different services.
| Requirement | Often the best starting point |
| Existing CUDA-dependent application or broad framework compatibility | NVIDIA H100, H200, B200, or B300 AI servers |
| Large-model inference requiring substantial GPU memory | H200, B200, or B300 server infrastructure |
| Evaluating an alternative AI accelerator for compatible workloads | Tenstorrent Wormhole AI servers |
| Dedicated capacity with predictable monthly cost | Dedicated NVIDIA or Tenstorrent AI hosting |
| Customer-owned, custom AI cluster | AI colocation in NYC or New Jersey |
| Private routing, carrier diversity, or a New York network edge | AI servers or colocation paired with Metanet connectivity |
This multi-platform approach is an advantage for a buyer. It allows the company to test a workload where it makes sense, keep the tools that already work, and avoid locking its entire product roadmap to a single hardware decision.
Galaxy, Wormhole, and newer Tenstorrent models
Galaxy belongs prominently in the conversation because it is a server-class Tenstorrent platform for customers that need more than a single accelerator. A Galaxy deployment is aimed at scalable AI compute: a serious platform that can be paired with high-capacity networking, storage, private connectivity, and a production operations model. It is the right discussion for AI providers that anticipate sustained inference demand, want a multi-accelerator environment, or need a path from initial deployment into a larger AI service.
Metanet can offer Galaxy systems as part of a broader data-center design. That includes rack placement, power and cooling review, out-of-band management, storage, Internet transit, private network access, and the operational support needed to run an AI hosting product rather than a lab machine. Galaxy therefore reinforces the core message: Tenstorrent hosting at Metanet can scale from available Wormhole capacity to a more substantial dedicated AI platform.
Wormhole is part of Tenstorrent’s evolving hardware roadmap. Customers evaluating a dedicated Wormhole AI server today should also understand that newer Tenstorrent platforms exist, including Blackhole products. Tenstorrent’s newer Blackhole line includes p100a, p150a, and rack-oriented p150b cards; the platform provides up to 32 GB of GDDR6 per processor and the p150 cards add four 800 Gb/s QSFP-DD connections for compatible multi-card configurations. Tenstorrent Blackhole documentation The p300c is a dual-Blackhole card with 64 GB of aggregate GDDR6 memory. Tenstorrent p300c specifications
That does not make existing Wormhole stock obsolete. Wormhole can be a highly useful AI hosting platform for supported workloads, especially when customers need capacity now and want to validate a model on an alternative accelerator. The presence of newer models gives customers a forward path: start with a proven Wormhole hosting deployment, then assess Blackhole or later-generation hardware as the application, budget, and available platforms evolve.
The practical advantage of working with Metanet is that we can have that discussion at the infrastructure level. We can help determine whether a current Wormhole server meets the immediate requirement, whether a newer platform should be planned for, and whether the customer’s application is better served by NVIDIA, Tenstorrent, or a mixed AI environment.
Why host Tenstorrent AI servers with Metanet?
An AI accelerator only becomes a usable product when it is deployed with the power, network, storage, security, and support behind it. Metanet brings data-center and ISP experience to the AI hosting conversation.
Customers can combine Tenstorrent AI servers with:
- Dedicated bare metal hosting.
- NYC and New Jersey AI data-center locations.
- Carrier-neutral connectivity and scalable Internet transit.
- Custom IP allocations and BGP routing for qualified customers.
- Private cross-connects and network architecture support.
- Remote hands for physical support and deployment assistance.
- AI colocation for customers adding their own infrastructure later.
- A growth path from one server to a more substantial AI deployment.
For some AI companies, the unique value of an NYC/NJ deployment is network proximity. 60 Hudson Street and 85 Tenth Avenue are important Manhattan connectivity environments, and NYIIX lists both as New York metro locations. NYIIX locations A customer does not need to colocate every compute node in Manhattan, but it can use the New York network edge for interconnection, carrier choice, and enterprise access while scaling other AI infrastructure in New Jersey.
Frequently asked questions about Tenstorrent AI hosting
Is Tenstorrent Wormhole a GPU?
Wormhole is an AI accelerator platform, not a conventional graphics GPU. It is designed for AI compute and data movement. For buyers, the relevant issue is whether the intended models and software stack are supported and perform well on the platform.
Can Wormhole run every CUDA workload?
No. CUDA applications are built for NVIDIA’s ecosystem. Tenstorrent should be evaluated with the customer’s actual models, frameworks, and operational requirements. Metanet’s role is to provide the dedicated hosting and network foundation; the model-fit review remains essential.
Is Tenstorrent appropriate for AI inference?
It can be a strong option for compatible inference workloads. The correct approach is to validate the model and desired throughput on the Wormhole platform before committing to a production deployment.
Can I begin with Wormhole and expand to Galaxy or newer models later?
Yes. A customer can start with available Wormhole AI servers, build and validate an AI service, then expand into Galaxy, evaluate Blackhole or later Tenstorrent models, or add NVIDIA and customer-owned colocated infrastructure as demand grows.
Build an AI hosting platform that gives you options
AI infrastructure should serve the product, not dictate it. Tenstorrent Wormhole servers, Galaxy systems, and newer models provide a real alternative for companies that want dedicated AI hosting, an open development path, and a platform they can test against their actual inference workloads.
Metanet offers available Tenstorrent Wormhole servers, Galaxy systems, other Tenstorrent configurations, dedicated AI hosting, NVIDIA H100/H200/B200/B300 infrastructure, and AI colocation in NYC and New Jersey. Contact Metanet to discuss your model, expected utilization, network design, and the right AI server platform for your deployment.
