Why I'm Bullish on Edge Compute (Again)

The First Wave (That Fizzled)

Remember 2018? Edge computing was going to revolutionize everything. IoT devices would process locally. CDNs would run full applications. Every startup pitch deck had "edge-first architecture" somewhere in the stack slide.

Then... nothing happened.

Not because the technology didn't work. It did. But the economics didn't. Cloudflare Workers and AWS Lambda@Edge were cool demos—but for most workloads, the complexity overhead crushed any latency gains. Centralized cloud was just easier.

So edge became a niche play: CDN static assets, maybe some A/B testing logic. The revolution got shelved.

What's Different Now?

Three forcing functions are making edge compute unavoidable in 2026:

1. AI Inference Economics

Running a 7B parameter model in a centralized data center costs $0.001 per request. Sounds cheap—until you're handling 10M requests per day. That's $10k/day, or $3.6M/year.

Move that inference to the edge and costs drop 70-90%. Why? You're not paying for:

• Round-trip latency to us-east-1
• Data egress fees (AWS's favorite profit center)
• Over-provisioned centralized GPU clusters

With AI models proliferating into every product, inference cost is becoming a first-order concern. Edge is the escape hatch.

2. The 100ms Rule

Real-time applications—multiplayer games, video calls, collaborative editing—have a hard ceiling: 100ms latency. Above that, the experience breaks. Physics doesn't negotiate.

If your user is in Singapore and your server is in Frankfurt, you're already at 160ms just for the round trip. Add TLS handshake, database query, API overhead—you're at 300-500ms. Dead on arrival.

The only solution? Compute closer to the user. Not "edge as a CDN." Edge as a full execution environment.

3. Data Sovereignty Is Real

GDPR was the warning shot. But in 2026, data localization laws are everywhere. China, India, Brazil, the EU—they all require local data processing for certain workloads.

You can't serve global customers from a single region anymore. Multi-region cloud gets expensive fast. Edge compute gives you local processing without managing 50 data centers.

The New Edge Stack

Here's what the winning architecture looks like:

Edge Layer (User-Facing)

• Cloudflare Workers / Deno Deploy / Fly.io — Run lightweight JS/WASM
• Vercel Edge Functions — Next.js-native, great DX
• Fastly Compute@Edge — High-performance, less hand-holding

Use for: Authentication, request routing, A/B testing, lightweight AI inference (small models < 1B params).

Regional Layer (Heavy Lifting)

• Kubernetes on bare metal (Hetzner, OVH, Vultr)
• Fly Machines (best UX for regional compute)

Use for: Database writes, large model inference, stateful workloads.

Core Layer (Source of Truth)

• Centralized database (Postgres with read replicas)
• Object storage (S3, Backblaze B2)
• Control plane (monitoring, config management, billing)

When Edge Still Doesn't Make Sense

Don't cargo-cult this. Edge is not a universal solution.

Skip edge if:

• Your workload is batch processing (no latency sensitivity)
• You're pre-product-market fit (optimize for iteration speed, not infra elegance)
• Your entire user base is in one geography
• You're running massive stateful workloads (databases, long-running jobs)

Edge shines when latency is a product feature and scale is predictable. If neither is true, centralized cloud is still the pragmatic default.

The Future Is Hybrid

The real insight: edge and cloud aren't competitors—they're layers.

Smart architectures use edge for latency-critical, stateless work, regional clusters for heavy compute, and centralized core for source-of-truth data. The traffic flows intelligently between them.

This is exactly how we're building DDoS mitigation at Link11: edge nodes detect and filter attacks, regional clusters run ML-based threat analysis, core systems handle billing and customer dashboards. Each layer does what it's best at.

Why This Time Is Different

The 2018 edge hype died because developer experience was terrible and cost savings were marginal.

In 2026:

• DX is solved: Deploying to Fly.io or Cloudflare Workers is easier than setting up AWS
• Economics are compelling: AI inference and data egress costs make edge a P&L decision, not a vanity metric
• Tooling matured: Distributed tracing, edge databases (Turso, Fly Postgres), and edge-native frameworks (Remix, SvelteKit) actually work now

The boring infrastructure improvements—the stuff that doesn't make headlines—finally made edge pragmatic.

The Bottom Line

If you're building anything user-facing with AI, real-time interaction, or global reach, you need an edge strategy. Not because it's trendy. Because the unit economics and user experience demand it.

The best ideas don't need permission. They need proximity to users. That's what edge gives you.

The revolution didn't die in 2018. It was just early. Now it's time.

---