Vercel Performance: Benchmarks, Latency & Limits 2026
OpenStatus clocked Vercel Edge at 106ms p50 against an 859ms cold Serverless start. Functions scale to 30,000 concurrent on a 99.99% Enterprise SLA.
Vercel Performance verdict
Vercel is edge-first.
126+ points of presence in 51 countries route to 20 AWS-backed compute regions over a private backbone. The independent numbers back the design.
Know two ceilings first. Concurrency tops out at 30k on Hobby and Pro, so size for that before you assume Enterprise. Cold Serverless starts near 859ms, real latency on the first hit. From there the choices are clear. Use Edge for latency-sensitive global routes that want p50 near 106ms. Run Serverless with Fluid compute for I/O-heavy work, and keep functions warm to dodge that cold start. Pin each function to the region nearest its database.
- The OpenStatus latency reflects 6 probe points hitting a function in iad1. Real TTFB depends on how close the visitor is to a point of presence and where the function runs.
- Cold-start figures apply to Node.js Serverless functions. The Edge runtime is consistently faster but runs a restricted API, with no Node built-ins and a 25s response-initiation deadline.
- Burst ramp is capped at 1,000 new executions per 10s per region, so absorbing a large spike takes minutes even though the ceiling reaches 30k to 100k. The 1800s extended duration is a per-function beta, not a project default.
- Edge p50 (independent)
- 106 ms
- Serverless cold p50
- 859 ms
- Max concurrency
- 100,000+
- Max duration
- 800 s
- Enterprise SLA
- 99.99%
This page covers how fast Vercel runs and how it scales under load. Region coverage and pricing live on their own pages.
Plan your Vercel concurrency and latency
- A peak of 5,000 fits the 30,000 auto-scale cap on Hobby / Pro (17% used).
- Serverless warm p50 sits near 246 ms at this load; a cold start runs to 859 ms, so push load higher and it drifts up.
- At 246 ms effective p50, that cap sustains roughly 121,951 req/s steady-state.
- Ramping cold to that peak takes about 50s at 1,000 new executions per 10s per region.
Estimated from Vercel's published concurrency caps and independently measured latency. Verify against your own workload.
Vercel query latency profile
| Runtime | Percentile | Latency | Sample |
|---|---|---|---|
| Edge | p50 | 106 ms | 6,042 pings, 6 regions |
| Edge | p75 | 124 ms | 6,042 pings, 6 regions |
| Edge | p99 | 328 ms | 6,042 pings, 6 regions |
| Serverless (warm) | p50 | 246 ms | 12,090 pings, 6 regions |
| Serverless (cold) | p50 | 859 ms | 2,010 pings, 30-min cadence |
| Serverless (cold) | p99 | 1,156 ms | 2,010 pings, 30-min cadence |
Vercel throughput and rate limits
| Limit | Ceiling | Notes |
|---|---|---|
| Max concurrency (Hobby / Pro) | 30,000 | Auto-scales per account based on demand |
| Max concurrency (Enterprise) | 100,000+ | Higher ceilings negotiable on Enterprise |
| Burst ramp rate | 1,000 / 10s / region | Initial increase per region before the next step |
| Throttle error on exceed | 503 FUNCTION_THROTTLED | Returned when the burst limit is exceeded |
| File descriptors | 1,024 shared | Shared across all concurrent executions including runtime |
| Edge response initiation | 25 s | Edge runtime must start responding within 25s to keep streaming |
Vercel scale ceilings
| Limit | Hobby | Pro / Enterprise |
|---|---|---|
| Max memory / vCPU | 2 GB / 1 vCPU | 4 GB / 2 vCPU |
| Configurable max duration | 300 s (fixed) | 800 s |
| Extended max duration (beta) | not available | 1800 s (30 min) |
| Bundle size (uncompressed) | 250 MB | 250 MB |
| Request / response payload | 4.5 MB | 4.5 MB |
| Default max duration | 300 s | 300 s |
Vercel reliability and architecture
- The Enterprise SLA guarantees 99.99% monthly uptime, calculated as Uptime / (Total minutes - Excused Downtime) x 100
- SLA credit tiers: 10% of monthly fees for uptime 99.1-99.98%, 25% for 95-99%, 50% for below 95%, capped at 50% of monthly fees
- Fluid compute is enabled by default for new projects since April 23, 2025, and lets multiple invocations share one function instance, which suits I/O-bound work like AI and database calls
- Bytecode caching on Node.js 20+ stores compiled JavaScript after first execution so later cold starts skip recompilation
- Availability-zone failover is automatic: on AZ failure, traffic moves to another AZ in the same region before failing over to the next region, for both fluid and non-fluid deployments
- Infrastructure runs primarily on AWS with data encrypted at rest using AES-256, and backups every two hours retained for 30 days
Vercel latency benchmarks, independently measured
- OpenStatus, a third-party monitoring service, benchmarked Vercel with 12,090+ Serverless and 6,042 Edge measurements across 6 global regions (Amsterdam, Sao Paulo, Hong Kong, Washington DC, Johannesburg, Sydney)
- Independently, Edge functions are about 9x faster than Serverless during cold starts and 2x faster when warm
- Independent Edge p50 of 106ms versus a cold Serverless p50 of 859ms confirms Edge for latency-sensitive global routes
- Vercel terminates TCP at 126+ PoPs across 51 countries and routes over a private low-latency backbone to the nearest of 20 compute regions
- Every deployment gets unmetered always-on DDoS mitigation and automatic TLS 1.2/1.3 at no extra cost
- Static assets are cached and served from the PoP nearest the visitor, while compute runs in the configured function region
Vercel Performance FAQ
How fast are Vercel Edge functions compared to Serverless?
OpenStatus measured this independently across 12,000+ requests from 6 global regions. Edge came in at 106ms p50, 124ms p75 and 328ms p99. Warm Serverless ran 246ms p50. Cold Serverless reached 859ms p50 and 1,156ms p99. Edge is roughly 9x faster on a cold start and about 2x faster warm, because it runs on the CDN layer and skips Node.js runtime initialization.
What is Fluid compute and how does it reduce cold starts?
Fluid compute lets several invocations share one running instance, so idle capacity soaks up requests without a per-request cold start. It has been the default for new projects since April 23, 2025. On Node.js 20 and up it also caches bytecode, storing compiled JavaScript after the first run so later cold starts skip recompilation. Billing counts active CPU time rather than I/O wait.
What are the function memory and duration limits per plan?
Hobby gives you 2 GB of RAM, 1 vCPU and a fixed 300s. Pro and Enterprise default to the same 2 GB and 1 vCPU, but scale to 4 GB and 2 vCPU. The maximum duration is 800s, or 1800s in beta for specific runtimes. The request payload is capped at 4.5 MB on every plan, and the deployment bundle at 250 MB uncompressed.
How does Vercel auto-scale functions under traffic spikes?
Functions auto-scale to 30,000 concurrent executions on Hobby and Pro, and past 100,000 on Enterprise. The limit that bites is the ramp. Vercel adds at most 1,000 new concurrent executions per 10 seconds per region, so a large spike takes minutes to fully absorb. Cross that burst limit and the request comes back as a 503 FUNCTION_THROTTLED error.
What SLA does Vercel offer and what credits apply?
The Enterprise SLA guarantees 99.99% monthly uptime, not counting excused downtime. Credits step down from there. You get 10% of monthly fees back for uptime between 99.1% and 99.98%, 25% between 95% and 99%, and 50% below 95%, capped at 50% of the monthly fee. The SLA does not cover the APIs or the CLI components.
Are Vercel's performance numbers independently verified?
The latency figures are. They come from OpenStatus, an independent monitoring service that published 12k+ cross-region measurements on its blog. The rest come from Vercel's own documentation: the concurrency caps, the memory and duration limits, the SLA credits. Those are platform limits rather than something an outside probe can measure.
Sources & verification
| Source | What was checked | Last checked |
|---|---|---|
| Vercel Official | Official product page | July 10, 2026 |
| OpenStatus | Independent latency benchmark | July 10, 2026 |
| Vercel Cdn | CDN and edge network | July 10, 2026 |
| Vercel Fluid Compute | Compute model | July 10, 2026 |
| Vercel Functions Concurrency Scaling | Concurrency and scaling | July 10, 2026 |
| Vercel Functions Limitations | Limits and quotas | July 10, 2026 |
| Vercel Security | Security and compliance | July 10, 2026 |
Every fact on this Vercel page is tied to a named source and a verification date. Freshness-sensitive figures trace to the sources above; verify against the vendor before relying on them.
Explore Vercel
Every page on Vercel in one place, you are on performance.