VoltageGPU vs Together AI

Together AI and VoltageGPU sit closer to each other on the product map than most of our other comparison pages. Both are OpenAI-compatible inference platforms — change the base_url from api.openai.com/v1 to api.together.xyz/v1 or api.voltagegpu.com/v1 and existing client code keeps working. Both serve open-weight models from the same families: Qwen 2.5 / 3.x, Llama 3.x, DeepSeek V3, gemma-class small models. Both bill per token with no minimum commit. From a developer-integration point of view the two platforms are interchangeable in a thirty-second swap of an environment variable.

The product-shape delta is what happens to a prompt after it crosses the API boundary. On Together AI the prompt enters a US multi-tenant inference stack operated by Together Computer, Inc. — a Delaware corporation headquartered in San Francisco — running on standard hypervisors with SOC 2 Type II controls. There is no Intel TDX VM around the inference worker, no GPU TEE on the H100/H200 silicon, no Intel DCAP attestation quote that an auditor can re-verify offline, and no European operator on the contract. That is sufficient for the majority of inference workloads and Together has built one of the strongest brands in the open-weight inference category on exactly that posture.

On VoltageGPU the same prompt enters an Intel TDX guest VM with AES-256 memory encryption, traverses an NVIDIA Protected PCIe link encrypted in hardware to a GPU sitting inside the trust boundary, and is processed by a model weight set loaded into TDX-sealed memory. Every confidential session exposes an Intel DCAP attestation endpoint that signs the entire configuration against the Intel root certificate; the cryptographic evidence that the operator could not read the prompt is delivered fresh for every call. The operator is VOLTAGE EI in Solaize, France, registered under SIREN 943 808 824, and the contract is signed against the French legal framework. That is the structural property the silicon enforces — not a marketing checkbox on top of an otherwise standard inference stack.

Together AI publishes a SOC 2 Type II attestation and operates a Data Processing Addendum for enterprise customers. That is the standard compliance posture every serious US inference platform exposes and it is sufficient for a large fraction of enterprise workloads — internal copilots, public-content summarisation, code generation against non-sensitive repositories, marketing-content pipelines, evaluation suites, and the long tail of AI work where the inputs were already non-sensitive when they reached the API. For those workloads Together is a price-competitive, broad-catalogue, fast-Turbo-tier inference platform and there is no good reason to pay more.

The posture stops being sufficient at the line where the technical measures clause of an Article 28 DPA needs to be backed by hardware evidence that the operator cannot read prompt memory. A SOC 2 audit confirms organisational controls; it does not constrain the host operator at the silicon layer. If a US administrator with legitimate console access wants to introspect a running inference worker — for legitimate operational reasons or under US legal compulsion through the Cloud Act framework — the standard hypervisor permits it. For prompts containing client files protected by French bar-association secrecy, health data under HDS scope, cardholder data under PCI DSS, or any personal data covered by GDPR Article 9, CNIL and equivalent European authorities have begun to require the technical measure be cryptographically enforced rather than contractually promised.

VoltageGPU is positioned at that requirement. The data physically does not leave European infrastructure, the operator is a French entity inside European jurisdiction, the encrypted memory key is ephemeral and per-VM, the attestation quote is signed by Intel and verifiable offline, and the GDPR Article 28 DPA is signed against the French legal framework. The cryptographic evidence that the operator is mathematically constrained from reading workload memory is produced fresh for every confidential inference session. The structural answer to "where does the prompt live and who can read it" is therefore enforced in silicon, not in policy. Together AI does not offer that posture today and has not announced TDX inference; that is not a Together bug, it is the inference category they chose to build.

It would be dishonest to write a comparison page against Together AI without acknowledging the categories where they are clearly the better tool. Catalogue breadth is the most obvious: Together serves 200+ open-source models including image generation (FLUX.1 dev, Stable Diffusion XL), video generation (Wan, Hunyuan), audio, embeddings, rerankers, and safety classifiers. VoltageGPU lists 16 TEE-attested text and code models — Qwen 3.x, gemma-class, DeepSeek-class, and frontier MoE variants — and does not currently ship a confidential image-or-video inference product. For any workload that needs FLUX, Wan, or Stable Diffusion at API speed, Together is the right answer and VoltageGPU is not in that category at all today.

On raw per-token economics Together wins outright on the tiny-model and frontier-output tiers. Llama 3.1 8B Instruct Turbo is $0.18 / $0.18 per million tokens on Together — we do not list an 8B TEE SKU, so for high-volume tiny-model inference where the workload is non-confidential, Together is the price-correct choice. More importantly, DeepSeek V3 is $1.25 / $1.25 flat on Together versus our Qwen3.5-397B-A17B-TEE frontier tier at $0.72 input but $4.33 output. The math is exact: on input we are 42% cheaper, on output Together is 3.5x cheaper. For an output-heavy frontier workload that does not need TEE — long-form generation, code synthesis at scale, agent-loop reasoning where output tokens dominate — Together wins the economic comparison and the buyer should choose Together.

Together also ships a dedicated-endpoint product starting at $3.99/hr for an H100 80GB that lets enterprise customers buy predictable throughput rather than burst per-token capacity. That is a category VoltageGPU has not built today on the inference side — our confidential GPU pods exist, but the "dedicated managed inference endpoint with auto-scaling" shape is a Together product feature we do not match. Turbo-tier latency on Llama 3.3 70B at 200–400 tokens per second is industry-leading and not a number we publish on our confidential serving infrastructure.

The honest summary is product-shape, not provider-quality: Together is the breadth-and-speed king on US multi-tenant open-weight inference and remains the right answer for unconstrained workloads on the broadest model catalogue. VoltageGPU is European confidential inference and is the right answer when the regulator, the client contract, or the threat model requires the prompt be hardware-sealed from the operator. A workload that picks the wrong-shape tool will be unhappy with both.