VoltageGPU vs Luminance

Most legal AI vendors are thin wrappers around OpenAI or Anthropic. Luminance is one of the few that actually built proprietary LLMs in-house — Luminance Diligence and Luminance Corporate AI are not OpenAI calls dressed up as a product. That is a real technical moat, it took a decade of work out of Cambridge, and it produces measurable benefits on legal-corpus accuracy that wrapper products cannot match. Any honest comparison page has to start there.

What that proprietary investment does not do is change where the models physically run. Luminance hosts on standard UK/EU cloud infrastructure. The models are theirs; the hypervisor underneath is not. From the workload's point of view this means Luminance the company retains technical read access to prompt content during inference — the legal contract restricts how that access is used, but the silicon does not. For most mid-market law firms working on routine contract review, that contractual posture is fine, and Luminance's model quality is the right reason to choose them.

The hardware-seal axis is independent of the model-quality axis. A best-in-class proprietary model can still leave the operator with read access to your client data, and a sealed open-weight model can be hardware-prevented from leaking that data even if the operator is hostile or subpoenaed. Those are different problems with different solutions, and Luminance is built to solve the first one. VoltageGPU is built to solve the second one. The two are not interchangeable, and pretending otherwise would be dishonest.

M&A diligence rooms see the most confidential corporate data inside any law-firm engagement. The full cap table sits in there. So do the term sheets, the unsigned letters of intent, the IP filings under non-public review, the employee compensation data, the customer contracts with confidentiality clauses, the unredacted financial statements, and the side-letters that the seller specifically did not want anyone outside the data room to see. The threat surface is wider than almost any other legal workflow because every category of sensitive data the firm handles ends up in the same place at the same time.

Intel TDX is the architectural answer to that threat surface. The CPU itself encrypts VM memory with an AES-256 key that the CPU generates internally and never exposes — not to the hypervisor, not to the operator, not to a system administrator with root on the host. The GPU side is sealed by NVIDIA Protected PCIe so model weights and tensor data crossing between CPU and GPU are encrypted on the bus. An Intel DCAP attestation quote signs the entire configuration with an Intel-root certificate that a verifier can re-check offline. For diligence in regulated jurisdictions — UK FCA scrutiny of financial-services M&A, EU MAR scrutiny of public-company transactions, French RGPD scrutiny of any cross-border deal touching personal data — this is no longer an optional architecture decision.

Luminance Diligence runs in the cloud regions Luminance operates and the operator (Luminance Technologies Ltd, UK) holds keys to the infrastructure. That is a contractual privacy posture, backed by SOC 2 and ISO 27001 certifications and a credible compliance team. VoltageGPU's Due Diligence Analyst agent runs inside an Intel TDX guest with attestation produced fresh for every session. The operator (VOLTAGE EI, France, SIREN 943 808 824) is mathematically constrained from reading workload memory because the encryption key for that workload is bound to the TDX VM lifecycle and is destroyed when the VM ends. Both products are credible at different regulatory tiers. The silicon path produces cryptographic evidence the contract path does not, and for a sufficiently sensitive diligence room, that evidence is the only thing the buyer, the seller, and the regulator will all accept.

Luminance wins on proprietary LLMs purpose-built for legal corpora. Their models were trained on legal language with a specific eye to M&A diligence, contract anomaly detection, and clause-level classification — and the resulting quality on legal-domain benchmarks is a legitimate technical moat that open-weight 32B or 235B models trained on general internet text do not match for free. Luminance also wins on multi-language depth (the diligence workflow handles cross-border deals in dozens of languages with first-class quality), on the autopilot autonomous review mode released in 2023 which has had real production hours behind it, and on UK/EU operator residency without US sub-processors — a credible posture for European clients that flatly refuse US Cloud Act exposure.

Luminance loses on price transparency. The public pricing page returns a 404 as of May 2026 and every commercial conversation routes through an RFP with annual-commit terms. Public reporting clusters real deals at $3,000+ per month at the entry tier and $50K to $200K+ per year at the enterprise tier — but a buyer cannot self-serve, cannot start in one afternoon, and cannot cancel monthly. Luminance also loses on hardware seal — there is no Intel TDX in their stack, no GPU TEE, no attestation chain rooted in Intel silicon, no operator-blind compute. And Luminance loses on extensibility for teams that want to bring their own agents or workflows: there is no BYOA pattern equivalent to VoltageGPU's OpenClaw / CrewAI integration where a customer can ship a custom agent against the same confidential infrastructure that runs our shipped agents.

The honest framing is this. For pure M&A diligence on proprietary corporate corpora where the model accuracy is the dominant factor and the regulatory posture of contractual confidentiality is sufficient, Luminance may be the better fit and we would say so to a prospect's face. For workflows where the threat model requires hardware-rooted attestation, where the procurement team needs transparent per-month pricing without an RFP, where the legal operator must be inside European jurisdiction with a per-session cryptographic proof the operator cannot read the data, and where the engineering team wants to extend the platform with custom agents through OpenClaw and CrewAI — VoltageGPU is architecturally the right answer. Many firms will run both, and that is a rational outcome.