How each report is made
The same four steps for every report and every new edition. Corrections become a new edition, never a silent edit.
Researched
From papers, vendor documentation, price lists, advisories and source code. Each number comes from a source we name; RBlock runs no benchmarks of its own for these.
Traced
Every figure is tied to a source, a data row or a formula shown in the report. Prices and rates come from one dated snapshot, so two reports never disagree.
Checked
A reviewer who did not write the report re-checks each key number against its source and recomputes the derived ones. Vendor claims stay labelled as vendor claims.
Signed off
RBlock reviews the voice, the judgement calls and anything said about named vendors. Legal topics go to an outside lawyer before release.
The reports
Showing all 20 reports.
-
Report 01: TEE vs MPC vs FHE for LLM inference
What confidential computing, multi-party computation and homomorphic encryption each protect, what privacy costs per token, how each has failed, and where to run it in the EU. Edition 2, checked against papers, advisories and price pages on 3 November 2026.
Free PDF from this page on 24 November 2026.
-
Report 02: Data partitioning for private inference: patterns and pitfalls
Which way of splitting an inference request keeps the operator out, what it costs per million tokens, what each split still leaks, how to make one safe, and what research must solve next. Edition 2, checked against papers, vendor documentation, price pages and attack disclosures on 3 November 2026.
-
Report 03: Confidential GPUs in practice: Hopper and Blackwell CC field guide
Where to rent a confidential GPU in the EU and at what price, a runbook per route, the throughput you lose by workload shape, the cost per million tokens against plain GPUs, a verifier policy mapped to published attacks, and who you still trust. Checked against NVIDIA documentation, provider docs and repositories, price pages, papers and advisories on 3 November 2026.
-
Report 04: The operator can't look: threat models for zero-access inference
Who can still read your prompts under each architecture, what confidential inference really costs, what has broken it since 2023, how to score a vendor, and what to ask. Checked against vendor threat models, security bulletins, audits, papers and price pages on 3 November 2026.
-
Report 05: FHE for transformers: how far are we really?
What fully homomorphic encryption can protect today and at what price; where the speed-ups came from; how it compares with TEEs and MPC under the same threat model; a dated hardware roadmap; when encrypted 8B inference could fit a budget; a pilot checklist and a ranked list of open problems. Checked against papers, standards pages, vendor announcements and shared prices on 3 November 2026.
-
Report 06: Inference engines compared: proprietary vs open stacks
Which engine to pick, what it costs per million tokens against an API, how safe each one is to expose, who stands behind it, and where in the EU to run sensitive work. Checked against engine docs, advisory databases and price pages on 3 November 2026.
Free PDF from this page on 17 November 2026.
-
Report 07: Open vs closed models: $ per task, not $ per token
Which models are good value, which provider to buy an open model from, and when renting or owning a GPU beats an API, with EU electricity, residency and confidential computing priced in.
Free PDF from this page on 17 November 2026.
-
Report 08: The local rig buyer's guide 2026–27
Which machine a team should buy for private, open-weight inference in the EU: how many people it serves, what it costs a year in your country, what fits, how fast it runs, where it can sit, when to buy and how to run it safely. Checked against EU price listings, Eurostat tariffs, vendor specifications, community benchmarks and engine source code on 3 November 2026.
-
Report 09: Quantisation without regret: accuracy loss by method and model size
Which file to run for ten widely used open models, what each quantisation level saves in memory and rent, what fails first and what a pre-ship test costs, what API endpoints actually serve, and whose quantised files to trust. 393 published results consolidated and checked on 3 November 2026.
-
Report 10: Hybrid local + private cloud inference: reference architectures
Seven ways to split work between your own hardware and a private cloud, with who must still be trusted, what each leaks, what it costs per month at your volume, the pinned components and fail-closed settings to build it, how it fails and how to test it. Checked against papers, vendor documentation, repositories and price pages on 3 November 2026.
-
Report 11: Privacy-preserving LLM routing: techniques and leakage risks
Who reads a prompt on its way through a gateway or router, which settings to change on 12 gateways, what each privacy layer costs per million requests, where EU-only routing is possible, and how self-hosted gateway software has failed. Checked against vendor documentation, API listings, source repositories, security advisories and papers on 3 November 2026.
-
Report 12: Model harness privacy watch-outs: agents, tools, memory and logs
Where ten agent frameworks, their observability integrations and the MCP SDKs send prompts, tool results, memories and keys by default; the configuration to ship instead; and what it leaves open. Every code claim is pinned to a file, line and commit at the releases of 3 November 2026.
-
Report 13: RAG privacy: embedding inversion, membership inference and vector DB risks
What leaks from a retrieval-augmented generation system, who can get it, and what to change before go-live: hardening settings and probes per vector store, ten managed services compared, an erasure runbook, each store's time to fix, and which defences to fund. Checked against papers, CVE records, source code, Helm charts and vendor pages on 3 November 2026.
-
Report 14: Prompt and output leakage via side channels in hosted inference
Which provider and which settings protect you from a network observer, another customer, your own tenants or the host; what each API and engine reveals; and what each fix costs. Checked against 33 published attacks, 19 providers' documentation, engine code at release tags and shared market prices on 3 November 2026.
-
Report 15: State of open-model inference performance, Q4 2026
What it costs to serve open models per million tokens, by GPU, engine and request shape; which published benchmarks to trust for each hardware tier; how fast audited results move; and what to run locally. Compiled from MLPerf, independent tests and community data, priced on 3 November 2026.
Free PDF from this page on 24 November 2026.
-
Report 16: Benchmarking private inference: an open methodology
Which private-inference technique can meet a latency budget today, what it costs, which threats it stops, and an open, versioned method for reporting results with their trust assumptions attached. Methodology v0.9, for review. Evidence and prices as of 3 November 2026.
-
Report 17: EU private and trustworthy AI research: key papers and groups of 2026
Where European groups lead and where they are thin across five top venues, whom to partner with or hire from, what code and companies you can build on, what attestation can and cannot prove, and where to apply before May 2027. Second edition, checked against venue lists, papers, repositories, vendor bulletins and funder pages on 3 November 2026.
-
Report 18: EU AI Act and GDPR meet inference architecture: what changes technically
Which EU duties apply to LLM inference in November 2026, when they bite for your profile, what each architecture adds in paperwork and cost, who enforces, and how to respond to an incident in Spain. Checked against the Official Journal, Commission, EDPB, Court of Justice and Spanish official texts on 3 November 2026. Not legal advice.
-
Report 19: Public-sector AI in Spain and Europe: procurement signals and sovereignty requirements
What Spanish public buyers tender when they buy AI, what a vendor must hold to bid, who wins, which contracts come up next and which Spanish and EU rules land when. Built from the Catalan open-data census of AI-titled contracts including minor ones, a random sample of TED notices, tender documents and published ENS certificates, checked on 3 November 2026.
-
Report 20: The private inference landscape 2026: companies, funding, open tools
Who builds private inference, which vendors an EU buyer can pilot, who funds the field and at what cheque size, which open tools carry maintainer risk, and where an EU founder can still build. 58 companies, 49 funding rounds, 24 scored open tools and 10 public audits, checked on 1–3 November 2026.
No report matches these filters.
How to get the reports
4 reports are released free on this page in November. The other 16 are written and independently checked; ask us for any of them and we send it to you.
Free releases
- 06Inference engines compared: proprietary vs open stacks
- 07Open vs closed models: $ per task, not $ per token
- 01TEE vs MPC vs FHE for LLM inference
- 15State of open-model inference performance, Q4 2026
Each one can be downloaded free from this page from its release date.
On request
- Email us with the report number, your name and your organisation. Each card has a button that starts the email for you.
- We reply with the report as a PDF, usually within two working days.
- We keep your email only to answer you. Nothing is added to a mailing list.
Found a number we got wrong? Tell us.
Corrections
Send a correctionName the report, the page and the source. Fixes go into a new edition.
Research collaborations
Discuss researchThreat models, benchmarks and private-inference prototypes.