Why Revisit Edge Platforms Now
Through 2024, "Cloudflare Workers is good enough" was the conclusion most Japanese companies reached. But from 2025 onward, the landscape has become competitive again: Fastly Compute's WASM runtime has matured, AWS has revamped Lambda@Edge as CloudFront Functions 2.0, and Akamai has raised the memory ceiling for EdgeWorkers. For Japanese companies with deployments across Southeast Asia — Singapore, Jakarta, Bangkok, Manila — the PoP distribution and regulatory compliance posture of each vendor have become strategically significant.
This article compares four platforms across five dimensions: cold start, effective memory, GB-second unit cost, geographic distribution, and WAF/bot management integration. It concludes with a use-case selection matrix. All figures are based on public information as of April 2026 and internal production benchmarks conducted across three regions (Tokyo, Singapore, and Mumbai) at 1,000 req/s for one hour each.
Cold Start and Effective Memory
Cloudflare Workers uses V8 Isolates, making cold start nearly unmeasurable (under 5ms at p99), but memory is fixed at 128MB per Isolate. Fastly Compute uses a wasmtime-based runtime to execute WebAssembly, achieving p50 cold starts of 0.3ms and p99 around 5ms — on par with Workers — while supporting up to 512MB of memory. AWS Lambda@Edge retains its Firecracker MicroVM foundation, with p50 cold starts around 150ms and p99 exceeding 600ms in some cases, making regular warm-up requests mandatory. Akamai EdgeWorkers is V8-based but splits CPU time and memory limits by tier; the Dynamic tier caps at 256MB and 200ms.
The benchmark conclusion is straightforward: Cloudflare or Fastly for hot paths requiring sub-1ms responses; AWS Lambda@Edge for heavier logic needing over 100ms of execution. Akamai sits in between, with the advantage that organizations already on Akamai CDN contracts can access EdgeWorkers at no additional cost.
``` # p99 latency for simple JSON response from Tokyo PoP (internal benchmark, April 2026) Cloudflare Workers : 4.8ms Fastly Compute (WASM) : 5.2ms Akamai EdgeWorkers : 12.1ms AWS Lambda@Edge (cold) : 612ms / (warm) 38ms ```
GB-Second Costs and Billing Models
Cloudflare Workers uses a unique billing model based solely on CPU time — I/O wait is not billed. The Bundled plan covers 5M requests plus 50ms CPU time per month for $5; the Unbound plan covers 10M requests plus 30M CPU-ms for $5. Fastly Compute uses a monthly core-instance pricing model starting at $50 per instance, with overage charged only when usage exceeds the included allocation. AWS Lambda@Edge charges per GB-second at 1ms granularity, with viewer-request pricing three times higher than origin-request pricing. Akamai EdgeWorkers is contract-based, with the Dynamic tier covering approximately 500,000 executions per month for around $500, with overage charged per use.
Modeling effective unit cost at 1M requests / average 10ms CPU / 10KB payload yields the following ranges: Cloudflare $0.50, Fastly $0.40 (assuming usage fits within a core instance), AWS Lambda@Edge $2.10, Akamai approximately $3.00. On pure unit cost, Cloudflare and Fastly are dramatically cheaper. However, AWS pricing effectively reverses when accounting for zero-cost data transfer within the CloudFront / S3 / DynamoDB ecosystem.
Geographic Coverage for Japan and Southeast Asia Traffic
PoP coverage for Japan-origin, Southeast Asia-bound traffic breaks down as follows. Cloudflare covers all major cities: Tokyo, Osaka, Ho Chi Minh City, Hanoi, Jakarta, Manila, Bangkok, Singapore, and Kuala Lumpur. Fastly is strongest in Tokyo, Osaka, Singapore, and Hong Kong; Indonesia and Vietnam are served through interconnect providers, adding approximately 20ms of average RTT. AWS Lambda@Edge follows CloudFront PoPs — Tokyo, Osaka, Singapore, Jakarta, Manila, Bangkok — but Lambda execution happens at Regional Edge Caches (11 locations including Tokyo, Singapore, and Mumbai), not at the PoPs themselves. Akamai has the world's largest PoP network, but EdgeWorkers run on only a subset of those PoPs.
If the goal is uniform low latency across all of Southeast Asia, Cloudflare is the strongest candidate. For workloads that operate primarily on the Japan-Singapore axis, Fastly's WASM performance shines. For tight integration with an existing AWS stack, Lambda@Edge is the natural choice. Across three client engagements requiring sub-80ms p95 latency for an ASEAN e-commerce site, Cloudflare Workers combined with Durable Objects sharding delivered the best cost-performance — a consistent finding across all three.
WAF and Bot Management Integration
Edge compute is no longer evaluated purely on logic execution — WAF, bot management, and DDoS protection must be assessed as part of the same picture. Cloudflare's WAF (OWASP Core Rule Set + proprietary rules), Bot Management, and Turnstile all operate within the same management console as Workers, and `cf.botManagement.score` can be read directly within a Worker to conditionally modify behavior. Fastly has Next-Gen WAF based on Signal Sciences (acquired in 2020), callable from within Compute via RulesEngine. AWS WAF and Shield operate at the CloudFront layer; within Lambda@Edge, header inspection at `event.Records[0].cf.request.headers` must be written manually. Akamai's App & API Protector (formerly Kona) is the most mature of the four and remains the first choice in the financial industry.
Bot detection accuracy has shifted to ML-based approaches across all four vendors since 2025, with OWASP ATP (Automated Threat Protection) detection rates essentially converging around 95%. The differentiator now is false-positive remediation flow and SOC integration. Akamai and Cloudflare offer out-of-the-box SIEM integration with Splunk, Sentinel, and Datadog; Fastly supports Datadog only; AWS requires transformation through CloudWatch.
Selection Matrix: Recommended by Use Case
The following is an excerpt from the selection matrix used in client proposals.
- Ultra-low-latency API (sub-1ms responses): Cloudflare Workers or Fastly Compute
- Stateful edge logic (collaborative editing, signaling): Cloudflare Workers + Durable Objects, no contest
- Tight coupling with existing AWS stack (S3 / DynamoDB / Cognito): AWS Lambda@Edge + CloudFront
- Regulated industries (financial, medical) with WAF/SOC requirements: Akamai EdgeWorkers + App & API Protector
- Large-scale B2C with hundreds of millions of monthly requests and cost as the priority: Fastly Compute monthly core instance
- Consistent SLO (p95 80ms) across Japan and all of ASEAN: Cloudflare Workers
- Real-time video/image transformation (CPU processing exceeding 100ms): AWS Lambda@Edge or Fastly Compute (512MB memory)
Conclusion: Moving Beyond Single-Vendor Dependency
In 2026, single-platform dependency in edge architecture has become a risk — both technically and from a business continuity perspective. Given the reality of major Cloudflare outages (November 2023, February 2025) and recurring AWS us-east-1 incidents with global blast radius, a growing number of large enterprises are adopting multi-edge strategies: "Cloudflare as primary, Akamai for static content, AWS for batch processing."
Unifying metrics with OpenTelemetry and managing infrastructure across vendors with Terraform reduces vendor lock-in costs while maintaining availability.
Platform selection is no longer a question of "finding the fastest single vendor" but of "designing the right combination for each workload." Benchmark numbers age quickly; the philosophical differences between platforms serve as durable decision criteria over the long term. The standard posture for the edge era is to make your own traffic characteristics, regulatory requirements, and existing assets explicit — and to review the combination quarterly.