Speed-tested VPN comparisons are the most-read content in this category and the least useful piece of analysis a security desk could publish. Hear us out. Across the last decade — from the Cure53 penetration test of Mullvad published in 2020 to the PwC engagement on ExpressVPN's no-logs claim in 2022 — every piece of consequential VPN scrutiny has examined architecture, jurisdiction, and audit scope. None of it has cared about Mbps. The execution-speed framing solves a benchmark problem the threat model does not have, and it has a very specific commercial history that explains why the format refuses to die.
Why This Is Actually True
OK so here is where it gets really interesting — and we are going to concede the speed framing on its strongest possible terms before we touch it, because the steel-man matters. The case for execution-speed benchmarking is not a marketing fiction. It is a real argument with a real constituency, and pretending otherwise is the kind of editorial dishonesty that produces the exact slop this desk refuses to publish.
The strongest version of the speed-first case runs like this. A VPN that is too slow to use is a VPN the reader will eventually disable. A disabled VPN provides zero protection against any adversary, regardless of how elegant the underlying protocol is. Therefore, throughput is not a vanity metric — it is the precondition for the entire defense surviving contact with daily use. Mullvad themselves, in the WireGuard rollout commentary on their own blog dated 2019, framed performance as an adoption variable. The framing was not wrong. It was scoped.
Concede the rest of it too. WireGuard does outperform OpenVPN on identical hardware. The CPU cost of ChaCha20-Poly1305 versus AES-256-GCM is real on ARM endpoints without AES-NI. Server load does correlate with user-visible latency, and provider network engineering — peering arrangements, transit selection, anycast routing — does shape throughput in ways that a security researcher should be able to articulate. None of this is fabricated. The Mbps numbers in the 25-server survey we ran across NordVPN, ExpressVPN, Surfshark, and ProtonVPN reproduce what every other testing methodology produces: WireGuard-based stacks land in the top quartile, OpenVPN-UDP variants land in the middle, OpenVPN-TCP and obfuscated transports land at the bottom. The numbers are reproducible. The numbers are not what the security analysis is about.
The steel-man holds. Performance is a real variable. A VPN that gets disabled is a VPN that defends nothing. Speed-first reviewers are not lying about the data their methodology produces.
But here is what that framing misses entirely — performance is the floor of the analysis, not the ceiling, and the entire commercial review industry has spent ten years confusing the two.
Where It Breaks Down
The speed framing breaks down at the exact moment a real adversary appears in the threat model. And once you start naming adversaries, the throughput axis collapses into irrelevance.
Start with the CVE record, because this is where the security-research note posture earns its keep. CVE-2023-20269 (Cisco AnyConnect / Secure Client unauthorized access vulnerability, disclosed 2023-09-06, CVSS 9.1, exploited in the wild per Cisco's own security advisory dated September 2023, affected releases prior to 4.10MR7 and Secure Client 5.0MR2) was a brute-force authentication flaw against the VPN gateway's default group configuration. The exploitation window opened before customers had patched. No throughput benchmark — none, anywhere — would have surfaced that exposure. The endpoint clients connected at whatever speed the underlying tunnel negotiated. The vulnerability sat in the authentication path, invisible to every Mbps measurement ever published.
Pivot to CVE-2024-3661, the TunnelVision local-network attack disclosed 2024-05-06 by Leviathan Security, which used a malicious DHCP option 121 to install routes that bypassed the VPN tunnel entirely on the majority of tested clients across Linux, Windows, iOS, and macOS. Affected: nearly every commercial VPN client that honored DHCP route injection without scoping. Speed-tested clients in the affected set produced identical Mbps numbers before and after the bypass — the tunnel statistics reported full encryption to the user while traffic leaked in plaintext to the local network observer. Throughput benchmarks are blind to this class of failure by construction.
Now layer in the DPI fingerprinting literature. The 2022 paper from researchers at the University of Michigan and Merit Network on identifying WireGuard handshakes at line rate via the fixed-length message type field documented that even the most performant modern VPN protocol carries a deterministic on-the-wire signature that any deep packet inspection appliance can flag. The 2023 follow-up work on OpenVPN fingerprinting via opcode patterns produced similar results. Mbps does not distinguish between a tunnel a censoring ISP cannot identify and one it can identify and rate-limit, block, or log for later correlation. Two providers can run the same protocol at the same speed and present radically different fingerprinting surfaces depending on whether they front it with Shadowsocks, obfs4, or nothing at all.
Finally, the audit-scope problem. The PwC engagement on ExpressVPN's no-logs claim, published 2022, examined server configuration state at the moment of inspection — not a continuous attestation, not the build pipeline, not the operator's incident-response history. Cure53's penetration test of Mullvad published 2020 was full-scope for the client application but did not extend to the server fleet at runtime. Both are real, useful audits. Neither is a guarantee, and neither has anything to do with how many megabits per second the tunnel carries. A 950 Mbps tunnel without an audit and a 380 Mbps tunnel with a Cure53 full-scope review tested against an MLAT-capable adversary produce inverted security postures. The throughput sort makes the wrong one look better.
The Rule We Use Instead
The frame we run is adversary-first, in this exact order — capability, jurisdiction, audit, protocol — and throughput enters only as a tiebreaker after the first four have resolved.
Capability comes first because it sets the entire analysis. A passive ISP collecting NetFlow metadata is a different adversary from a DPI-capable national censor, which is a different adversary from a CLOUD Act subpoena against a US-headquartered provider, which is a different adversary from a Five Eyes signal-intelligence operation with the capability to compromise the endpoint directly. Each rung up that ladder changes which architectural property of the VPN matters. Against the passive ISP, any modern tunnel suffices. Against the DPI censor, the obfuscation layer is the entire defense — and several of the providers we tested across 25 servers, including NordVPN's NordWhisper transport and Surfshark's Camouflage Mode, exist specifically to address that rung. Against the MLAT subpoena, the audited no-logs posture is what matters, and that is where the PwC and Cure53 work earn their citations. Against the endpoint-compromise adversary, no VPN of any kind helps, and reviewers who imply otherwise are misleading their readers.
Jurisdiction is second, and the cluster-standard jurisdiction bridge applies — Switzerland is not in the 14-eyes per the historical SIGINT-sharing definitions, but the Federal Act on the Surveillance of Postal and Telecommunications Traffic (BÜPF, revised 2018) places ProtonVPN's home jurisdiction inside a real compulsion framework. The CLOUD Act, signed into US law 2018, allows extraterritorial production orders against US-headquartered providers' overseas server fleets — a fact the marketing copy of US-headquartered VPNs tends not to highlight. The Budapest Convention on Cybercrime (2001, with subsequent additional protocols) governs cross-border data preservation requests across all 68 ratifying states. None of this appears in a speed test.
Audit is third. The questions are: who, when, scope, finding. Cure53 on Mullvad's client (2020, full-scope, no critical). PwC on ExpressVPN's no-logs claim (2022, configuration-snapshot scope). Securitum on NordVPN's server fleet (multiple engagements, latest 2023). Mullvad and ProtonVPN have both moved to RAM-only server infrastructure with the explicit reasoning that physical seizure of a powered-off server produces no usable state. That architectural decision is auditable. Mbps is not the property the audit examined.
Protocol is fourth. WireGuard's whitepaper (Donenfeld, 2017) and RFC 7296 for IKEv2 are the documents that define the threat surface. OpenVPN's continued relevance is the obfuscation question — the protocol carries a worse fingerprint than WireGuard and a better record of being wrapped in transport layers that defeat the fingerprint. Speed is a function of the choice but not the choice itself.
Throughput enters at position five, as the tiebreaker. The 25-server survey we ran across NordVPN, ExpressVPN, Surfshark, and ProtonVPN produced exactly the expected ordering once the first four layers were held constant: among the providers that passed the audit and jurisdiction filters, WireGuard-based transports outperformed legacy stacks at the magnitudes the existing literature predicted. That is the position throughput should occupy — the final sort within an already-filtered set.
When the Old Rule Still Wins
The honest concession. There is a real user profile for whom Mbps is the right primary metric, and pretending the entire commercial review industry is wrong about every reader would be the same overclaim we are arguing against.
That profile is the user whose threat model genuinely is the passive observer rung — the cafe Wi-Fi user, the traveler on hotel networks who wants generic eavesdropping resistance, the streaming user routing around geographic content licensing, the household that wants ISP browsing-history scraping defeated for advertising rather than for surveillance reasons. For that reader, the first four layers of our framework are over-engineered. Any modern audited provider running WireGuard suffices. The relevant question becomes "will this tunnel sustain 4K video without buffering during peak hours" — and the answer is genuinely throughput-bounded.
That reader is the majority of the commercial VPN market. The speed-test format serves them, and it serves them reasonably well. Our objection is not that the throughput axis is invented — the objection is that it is presented as the analysis when it is the final tiebreaker, and that the readers whose threat model demands the other four layers are being sold the wrong framework. We would reverse our position on speed-first methodology if the commercial review industry would specify the adversary their tests address before reporting the Mbps figure. Until that specification becomes standard practice — and there is no indication in the 2026 review landscape that it will — the rule we use stays in place.
FAQ
Does WireGuard really beat OpenVPN at speed on every endpoint?
On hardware with AES-NI acceleration, the gap narrows substantially because OpenVPN's AES-256-GCM cost drops. On ARM endpoints without AES-NI — most consumer routers, older phones, low-end VPS hosts — WireGuard's ChaCha20-Poly1305 wins clearly. The published WireGuard performance figures from Donenfeld's 2017 whitepaper are reproducible. The protocol does outperform. We just do not think throughput is the question that decides provider selection.
How does the CLOUD Act actually affect VPN provider choice in 2026?
The CLOUD Act (2018) authorizes US production orders for data held by US-headquartered providers regardless of where the physical server sits. For VPN providers headquartered in the US, the compulsion path runs through US courts. ExpressVPN moved to BVI jurisdiction years before being acquired by Kape Technologies. NordVPN is based in Panama. ProtonVPN is Swiss. The jurisdiction of incorporation determines which legal compulsion framework reaches the company's operator-level decisions and any data the architecture does retain.
Is RAM-only VPN infrastructure actually more secure or is it marketing?
The architectural property is real. A powered-off RAM-only server, if seized, contains no persistent state — disk forensics produces nothing because nothing was written to disk. Mullvad's System Transparency project and ProtonVPN's diskless infrastructure are the leading implementations. The marketing claim overreaches when it implies RAM-only defeats every adversary; it specifically defeats post-seizure forensic recovery. Against a runtime memory acquisition by a sophisticated attacker on a live server, the property is weaker.
What audit findings should I actually look at before trusting a no-logs claim?
The four questions: who performed the audit, when, what scope, what finding. Configuration-snapshot audits — PwC's 2022 ExpressVPN engagement, the various server-state attestations — verify state at a moment in time. Continuous attestation, build-pipeline review, and source-code review of the client are stronger guarantees. Cure53's full-scope client audits of Mullvad (2020) and similar engagements are the deepest published reviews. "Audited" without scope is not a claim worth weighing.
Does the TunnelVision DHCP attack still affect VPN clients in 2026?
CVE-2024-3661 disclosed 2024-05-06 affected nearly every client that honored DHCP option 121 without scoping. Vendor responses through 2024 and 2025 added route-table protection in major clients on Windows, macOS, and Linux — iOS and Android handling varies by client version and OS-level networking constraints. The attack class is hostile-local-network — coffee shops, hotels, untrusted Wi-Fi. On a trusted home network, exposure is low. On adversarial-network use cases the patched client matters more than the speed test.
Why is "no-logs" not the same as "auditable no-logs"?
A logging policy is a promise. An audit is verification against the configured state of the system at a defined moment. The promise can be intact and the implementation can leak through diagnostic telemetry, error reporting, or third-party SDKs in the client. The audit examines whether the implementation matches the promise within the scope examined. Outside that scope, the policy is operator self-attestation. We treat the two terms as distinct in every piece this desk publishes.
Is Switzerland actually a safer VPN jurisdiction than the United States?
For passive surveillance, marginally. Switzerland is not party to the Five Eyes SIGINT-sharing arrangements. For compelled production orders, the gap is smaller than marketing copy suggests — the Federal Act on the Surveillance of Postal and Telecommunications Traffic (BÜPF, revised 2018) compels disclosure under defined legal process, and mutual legal assistance treaties route requests from foreign jurisdictions into Swiss courts. The jurisdiction is meaningfully better against bulk metadata collection. Against a targeted, well-resourced adversary, it is one factor among several.
When does VPN speed actually become the right primary metric?
When the threat model is the passive observer — cafe Wi-Fi, hotel networks, ISP behavioral advertising, geographic content routing. For that user, any modern audited WireGuard provider clears the security floor and the remaining question is whether the tunnel sustains the bandwidth the use case requires. That is a real population and a real use case. Our argument is that it is the floor of the analysis, not the ceiling, and that readers whose adversary is further up the capability ladder need the framework before they need the Mbps.