Cloudflare
Cloudflare

The Platform Every
AI Workload Needs

GovTech × Cloudflare  ·  March 2026

Every Decade, a New Layer Rewrites the Rules

2000s

2010s

now

// web era · 2000s

HTTP

The browser was the platform

Delivery won. CDN, edge caching, DDoS protection. The layer between users and content defined who dominated the web.

infrastructure winner: cloudflare

// api era · 2010s

REST

The API was the platform

Connectivity won. API gateways, identity, payments. Building blocks that composed into products. New infrastructure winners emerged.

infrastructure winners: stripe · twilio · auth0

// agent era · now

MCP

The agent is the platform

Governance wins. Edge compute, orchestration, and security by default. The layer between the model and the real world is the new battleground.

cloudflare's conviction: we are building this layer

Our Bet

You bring the model.
We provide everything else.

// not a fourth cloud to manage alongside AWS, Azure, and GCP

// not a CDN that added AI features as an afterthought

// not a point solution that solves one workload and ignores the rest

The programmable layer your AI workloads run on. Sitting above all three clouds, with governance built in by default, and compute in Singapore.

The Platform

COMPUTE

Workers

330+ PoPs. Zero cold starts. CPU-time billing only.

Durable Objects

Agent memory and session state. Strongly consistent.

Workflows

Durable pipelines. Retries. Guaranteed execution.

Workers for Platforms

Isolated sandbox per agency or tenant.

AI

Workers AI

50+ open-source models. Serverless GPU. Pay per Neuron.

AI Gateway

Every LLM call: logged, DLP-scanned, cached, cost-attributed.

Vectorize

Vector DB for RAG and semantic search.

Agents SDK

Stateful agents with MCP client and server support.

DATA

R2

S3-compatible object storage. Zero egress fees. Data stays.

D1

Serverless SQLite at the edge. No connection pooling.

Workers KV

Globally replicated key-value. Sub-ms reads worldwide.

Queues

Async messaging. Decouple producer from consumer.

CONTROL

Zero Trust

Identity-aware access. Already deployed. AI inherits it.

MCP Gateway

Control which tools agents can call. Log every invocation.

AI Security

Prompt injection detection. DLP on every AI call.

Observability

100% log capture. No sampling. No Datadog bill.

// connected via Worker Bindings · in-process calls · no HTTP overhead · no egress between primitives

Where This Is Going

// build

Agents that act, not just respond

Workers + Agents SDK + MCP gives every agency a path to AI agents that schedule tasks, call internal APIs, maintain context across sessions, and hand off to humans when needed. On government-grade infrastructure, in Singapore.

Workers Agents SDK Durable Objects

// secure

Every AI call, governed before it lands

AI Gateway + AI Security for Apps means no LLM call crosses the network without DLP scanning, classification enforcement, and a complete audit trail. Prompt injection detection. Custom topic blocking. Available now, across any provider.

AI Gateway AI Security MCP Portal

// MCP is becoming the protocol of the agentic era. Like HTTP was for the web. cloudflare is where you build, host, and govern MCP servers.

In Practice: PAIR at 150K Officers

01

Officer queries PAIR. Hits AI Gateway first. Not the LLM provider.

Workers
02

DLP scans the prompt for Restricted/Confidential content. Non-compliant calls blocked before any external model sees them.

AI Gateway
03

Cache hit or fallback. Cached response returned if available; otherwise routed to best live provider, automatic failover.

AI Gateway
04

Response logged by ministry. Cost attributed. Full audit trail retained. Delivered to officer.

AI Gateway
pair-gateway.ts
// One URL change. Nothing else in PAIR changes.

const AI_GW =
  'https://gateway.ai.cloudflare.com/v1/' +
  '{account_id}/govtech-pair-prod/openai';

const res = await fetch(AI_GW + '/chat', {
  method: 'POST',
  headers: {
    'cf-aig-metadata': JSON.stringify({
      ministry: req.headers.get('x-ministry-id')
    })
  },
  body: JSON.stringify({
    model: 'gpt-4o',
    messages: [{ role: 'user', content: prompt }]
  })
});
// DLP · caching · fallback: gateway config

For GovTech's World

// scenario_01

PAIR governance across 16 ministries

PAIR logs are scattered across providers today. AI Gateway puts a single control plane in front of every LLM call: DLP enforcement, per-ministry cost attribution, and provider fallback, without touching the PAIR application code.

AI Gateway Workers AI

// scenario_02

Agency MCP portal: managed, not hand-built

Your engineering team built a custom MCP interceptor. Workers + MCP Portal replaces it with managed infrastructure: agencies publish internal APIs as MCP servers, the portal composes them behind a single authenticated gateway with DLP tag enforcement built in.

Workers MCP Portal AI Gateway

// scenario_03

Isolated AI agents across 700+ Singpass integrations

Workers for Platforms gives each agency integration a dedicated V8 isolate: its own access controls, KV storage, and audit trail. A healthcare AI agent cannot reach CPF data by design. Isolation is structural. Not a policy rule that can be misconfigured.

Workers for Platforms D1

The GovTech AI Stack

// built on cloudflare primitives · inherits your existing security posture

// users

Citizens  ·  150K+ Public Officers  ·  AI Tools (MCP clients)

// cloudflare network · already deployed · already trusted

CDN  ·  WAF  ·  DDoS  ·  Bot Management  ·  Zero Trust

// edge compute

Workers

routing · auth · logic

Agents SDK + DO

state · orchestration

Workers for Platforms

per-agency sandboxes

// ai gateway · every ai call governed before it reaches a model

logged· dlp· cached· cost-controlled· model fallback· audit trail

Workers AI / External LLMs

Llama · Mistral · OpenAI · Anthropic

Vectorize + Durable Objects

semantic memory · agent state

R2

object storage · zero egress · SG region

D1

serverless SQL · edge-replicated

Zero Trust · MCP Gateway

identity · tool governance

Who's Done This

// government · new york city

New York City Housing Authority

Stood up a citizen-facing digital service for 1.5 million users in under 5 months, with a team of fewer than 10. 250,000 cyber attacks blocked in the first 9 hours. Zero downtime. 600,000+ applications processed.

// government · apac

Taiwan National Medical Institutions

Unified security posture across national hospitals and research centres, with strict data sovereignty requirements enforced via geographic routing. 2.65 million attacks mitigated in a single day. Zero downtime.

// developer platform · apac · 260m users

Canva

260 million monthly users, sub-50ms latency globally, on Workers. Zero Trust access to 90+ internal apps across 5,000 engineers. Acquisitions secured and integrated in days using Cloudflare as the security baseline from day one.

// developer platform + workers ai

VSCO

$400K annual cloud savings by migrating to R2. 6x improvement in worst-market API latency. Workers AI powers a real-time collaborative AI canvas. Inference at the edge, no GPU provisioning required.

Where We Go From Here

01

AI Gateway on PAIR

Point PAIR's LLM calls through AI Gateway. See every call, cost by ministry, and any classified content that should not be leaving the network. No PAIR code changes required.

time_to_value: < 1 week

02

Workers + MCP Portal proof of concept

Host one agency's internal APIs as an MCP server on Workers. Run it alongside your existing custom interceptor. Compare governance coverage, developer experience, and engineering overhead.

time_to_value: < 2 weeks

03

AI Security discovery

Run AI Security for Apps' free endpoint discovery across GovTech's public web properties. Find every AI-powered endpoint, including the ones your security team does not know about yet.

time_to_value: < 1 day

Over to You

// the_governance_gap

Of PAIR, MAESTRO, and your agentic SOC work: where is the governance gap feeling most acute right now?

// build_vs_platform

Your team engineered a custom MCP interceptor. What would a managed platform need to do to earn that trust?

// data_sovereignty

Which AI workloads are blocked today by the requirement to keep data in Singapore, and what would unblock them?

// the_one_thing

If there is one AI infrastructure problem GovTech would solve this quarter, what is it?