How I would run discovery, PoCs and APAC-hours support for OpenRouter in Asia Pacific, built on OpenRouter's live API, plus a working desk that turns a discovery call into a residency-aware routing plan and explains what any request will do.
The first go-to-market hire outside the US. Pre-sales engineer beside the AEs, plus onboarding, account health and first-line API debugging in APAC hours, with clean handoffs to US Support.
"Where do our prompts go?" In-region routing covers the EU and US. In APAC the answer is built from provider pinning, ZDR, BYOK and workspace guardrails, and the CE has to explain each one precisely.
A four-tab desk on OpenRouter's live API: discovery notes, a residency-aware routing plan with code, a request simulator with error decoder, and an APAC coverage view.
| Takeaway | What it means for the APAC Customer Engineer |
|---|---|
| APAC demand is already large | Singapore billed 9.21% of tokens in the State of AI study, second to the US; Asia's share of spend rose from about 13% to about 31% over the study year. Simplified Chinese is the second prompt language at 4.95%. |
| Residency is the first technical objection | Regulated buyers in Singapore, Japan, Korea and India ask where prompts are processed. The honest answer names what is fail-closed (EU, US base URLs) and what is configuration (APAC pinning). |
| Most tickets are routing filters | A 503 that never clears, tool calls that work "sometimes", empty replies: each traces to provider preferences meeting a model's live endpoint list. The desk shows that list. |
| The seat writes the playbook | One regional person, US teams asleep. Every repeat question becomes a doc, a runbook step or a product note with deal evidence. |
One OpenAI-compatible API in front of every major lab and host, with routing, fallbacks and one bill. Company figures as published; API counts read on 10 October 2026.
| Item | Fact |
|---|---|
| Scale | "10+ trillion tokens per day" and "over 10 million developers and companies" (August 2026). Live API: 458 text models (659 across all output types), 112 providers, 1,375 text endpoints. |
| Ownership | Series B of US$113m led by CapitalG (May 2026). Joining Stripe, announced 19 August 2026, with "same name, same product, same roadmap" and a neutrality commitment. |
| Plans | Standard: 5.5% fee on credit purchases. Business: 8%, adds EU and US in-region routing and 1,000 workspaces. Enterprise: contract, invoicing, SSO and SCIM, SLAs, discounted fees. Inference is billed at the provider's list price on every plan. |
| Enterprise controls | Workspaces, guardrails (budgets, PII and secret detection, allowed data regions), ZDR, presets, BYOK for Bedrock, Azure and Vertex, Broadcast traces to observability tools, Security Center. |
| Trust | SOC 2 Type 2 on the trust centre; subprocessor list published; DPA signed on request from Business. |
| Payments | Cards, AliPay and USDC; USD billing. |
Demand from the State of AI study (billing location, November 2024 to November 2025). Supply from /api/v1/providers, which lists each provider's declared datacenter countries.
| APAC market | Share of tokens |
|---|---|
| Singapore | 9.21% (#2 after the US at 47.17%) |
| China | 6.01% |
| Korea | 2.88% |
| Japan | 1.77% |
| India | 1.62% |
| Asia, all | 28.61% (from about 13% to about 31% across the year) |
| Provider | Datacenters | Serves |
|---|---|---|
| Alibaba | CN, SG | Qwen family, 43 endpoints |
| Z.AI | SG | GLM family, 14 endpoints |
| DekaLLM | ID | gpt-oss, Gemma and others, 10 |
| Tencent | SG | Hunyuan, 5 |
| Moonshot AI | SG | Kimi, 4 |
| Nex AGI | SG | Nex, 2 |
Read: APAC-located supply today is the Asian labs' own first-party hosting. No Bedrock or Vertex endpoint in an Asian region appears in the list, and 81% of endpoints declare no datacenter, so "unknown" covers most of the catalogue.
What I would put in front of an APAC buyer, with the strength of each answer stated. The legal fit is the customer's counsel's call; the CE's job is an exact description of the data path.
| Customer asks | OpenRouter answer today | Strength | Shown in demo |
|---|---|---|---|
| "No provider may keep our prompts" | zdr: true per request or account-wide; 930 ZDR endpoints. Check fallback depth per model, since ZDR narrows the pool. | Enforced | Plan, Simulator |
| "Process in Singapore" | provider.only to Moonshot AI, Z.AI, Tencent, Nex AGI, plus a workspace guardrail so every key follows it. Models: Kimi, GLM, Hunyuan, Nex. | Configured | Plan: Singapore only |
| "Process in APAC" | Adds Alibaba (CN, SG) and DekaLLM (ID): 58 models in total, including gpt-oss and Gemma. | Configured | Plan: APAC |
| "Claude or GPT, in Singapore" | BYOK to the customer's own Bedrock or Azure deployment in an APAC region, with OpenRouter as the routing and billing layer. The residency claim rests on their cloud contract. | Customer-owned | Discovery notes |
| "Fail closed in region, like the EU" | Available for eu.openrouter.ai and us.openrouter.ai on Business and Enterprise. No APAC base URL yet: product feedback with deal evidence. | Not in APAC | Coverage tab |
| Market | Rule to raise in discovery |
|---|---|
| Singapore | PDPA Transfer Limitation Obligation; MAS Technology Risk Management guidelines for financial institutions. |
| Malaysia | PDPA as amended in 2024: cross-border transfer on conditions, whitelist removed. |
| Indonesia | PDP Law 27/2022; GR 71/2019 localisation for public-scope systems. |
| Japan, Korea | APPI and PIPA: consent or an equivalence basis for overseas transfer; public and finance buyers often expect a domestic region. |
| India, Australia | DPDP Act 2023 and sector regulators (RBI); Privacy Act APP 8 accountability for overseas recipients. |
Who an APAC buyer compares OpenRouter with, and the honest line for each.
| Alternative | Their edge | Gap vs OpenRouter | How I would position |
|---|---|---|---|
| AWS Bedrock, Azure AI Foundry, Vertex | APAC regions, existing commits, IAM | One cloud's catalogue; per-region model gaps; no cross-cloud failover | Keep the commit: BYOK their Bedrock or Azure keys, add the rest of the market and fallbacks on top. |
| LiteLLM (self-hosted) | Runs in their VPC, strong residency story | They operate it; no pooled capacity, no one bill, no live provider uptime data | OpenRouter as the upstream for LiteLLM, or replace the ops burden; compare engineer-hours. |
| Portkey (now Palo Alto Networks) | Governance and security buyer | Mainly a BYOK control plane; roadmap tied to a security suite | Lead with supply and failover; guardrails and Security Center cover the governance asks. |
| Vercel AI Gateway, Cloudflare AI Gateway | Zero markup, edge caching, platform fit | Smaller provider pools per model; no public per-endpoint price and uptime market | Fine for an app on that platform; OpenRouter for multi-team, multi-provider routing. |
| Alibaba Model Studio, BytePlus ModelArk | Singapore regions, local sales teams | Centred on their own models | OpenRouter already routes to Alibaba first-party and adds the other labs with one key. |
| Together, Fireworks, DeepInfra, SiliconFlow | Own GPUs, low open-weight prices | One supplier's capacity | They are OpenRouter suppliers; the router spreads load across them and fails over. |
Points a customer can trip on, each checked against the primary page on 10 October 2026. Small docs items, worth a line in a support macro.
| Topic | What the pages say | What I would tell a customer |
|---|---|---|
| Platform fee | The docs FAQ lists the 5.5% card fee and 5% crypto fee; the pricing page adds the Business plan at 8%. The FAQ's "no volume discounts" line predates the enterprise fee discounts on the pricing page. | Quote the pricing page; it is the current source. |
| Model list price | /api/v1/models shows one price per model, the cheapest endpoint. GLM 5.3 shows $0.039/M input there; Z.AI's own endpoint is $1.40/M. | Compare a lab's price with that lab's endpoint in /endpoints. |
| Promotional prices | Endpoint prices are net of pricing.discount; 23 providers ran one on 10 October. | Budget at list price: charged / (1 − discount). |
| Region names | Guardrail value europe; models filter region=eu. The in-region doc flags this itself. | A likely 403 cause after a guardrail edit. |
| Metadata header | The errors page shows X-OpenRouter-Experimental-Metadata in its guardrail example; the router-metadata page calls that the legacy name, still accepted. | Use X-OpenRouter-Metadata. |
| JD duty | How I would do it | Detailed in |
|---|---|---|
| Technical discovery: architecture, models, residency, compliance | A fixed question set (workload, volume, residency rule per market, retention, current stack, billing, security review) that maps each answer to an OpenRouter control. | Demo tab 1, §03 |
| Custom demos and PoC environments | Two-week PoC: workspace with guardrail, three candidate models on 200 real prompts, TTFT measured from their region, one forced provider failure, written readout. | §06 |
| Partner with AEs through close; proposals, QBRs, executive briefings | A solution brief per deal (requirement, config, monthly cost, risks, open questions) the AE can send; technical answers in proposals within one APAC business day. | Demo tab 2 |
| Objections, competitive positioning, SLAs | Answers per competitor and per residency ask, with strength stated; SLA terms routed through Enterprise. | §03, §04 |
| Product and engineering feedback | Monthly APAC note: asks ranked by deals and revenue, each with the request and the gap (an APAC base URL first). | §08 |
| Onboarding, adoption, account health | Go-live checklist per account; weekly look at error rates, fallback counts and spend per workspace; flag risks to the US CSM. | §06 |
| APAC-hours support, code-level debugging, handoffs | Reproduce with a minimal curl, run the request through the routing rules, decode the error, hand over with generation id and repro. | §07, demo tab 3 |
| Stage | What I do | Artefact |
|---|---|---|
| Discovery | Workload, models in use, monthly tokens, latency target, residency and retention rules, existing cloud commits, buyer and security reviewer. | Discovery notes (demo tab 1) |
| Design | Candidate models under the policy, priced at their volume; provider preferences; fallback models from other families; guardrail settings. | Routing plan and solution brief (tab 2) |
| PoC | Agree pass marks first. Their prompts, their rubric, TTFT from their region, one forced failure to show fallback. | Scorecard and readout |
| Security review | Trust-centre package, subprocessors, DPA, data path diagram, before the PoC ends. | Data path one-pager |
| Go-live | Keys per workspace, budgets, Broadcast to their observability tool, alerts, named escalation path. | Go-live checklist |
| Adoption | Error and fallback rates, spend by workspace, new-model opportunities; QBR with numbers. | Account health note |
Five tickets I expect in the first month, each loaded as an example in the demo's simulator with live endpoints.
| Customer says | Root cause | Fix | Owner |
|---|---|---|---|
| "503 on every call since we turned on ZDR" | The model's only hosts are outside the ZDR list, so no endpoint is left (Qwen 3.7 Plus has 2 endpoints, both non-ZDR). | Pick a model with ZDR endpoints, or move ZDR to fallback models in models: [...]. | CE closes it |
| "Tool calls work sometimes" | require_parameters is off, so hosts that do not list tools stay eligible (4 of 12 for DeepSeek V3.2). | Set provider.require_parameters: true. | CE closes it |
| "Empty replies, finish_reason length" | Reasoning model spent max_tokens on reasoning; docs say a retry does not help. | Raise max_tokens or cap reasoning effort. | CE closes it |
| "We turned fallbacks off, other providers still answer" | Typo allow_fallback; unknown keys are ignored. | Correct the key; add a schema check in their client. | CE closes it |
| "Stream returned 200, then an error" | Provider failed after headers were sent; failover stops once tokens stream. | Read the SSE error and error_type; retry policy on finish_reason: error. | Handoff with generation id if one host repeats |
A two-page APAC data-path document per market, reviewed by legal and security, that every AE can send. It shortens the first security review in each deal.
Singapore already bills 9.21% of tokens. Start with existing self-serve accounts in Singapore spending enough to need invoicing, SSO or guardrails: expansion before new logos.
Log every deal that stalls on residency with market, model and revenue. That record is the input product needs for an APAC in-region decision.
Most routing tickets close without engineering. A simulator-style check and saved macros let APAC customers get an answer the same day.
81% of endpoints declare no datacenter. Working with Provider Partnerships to fill it turns APAC pinning into a computed list.
Simplified Chinese is the second prompt language. Quickstarts and a discovery deck in 中文 for Singapore, Malaysia and Taiwan teams.
Shadow AE calls and the US Support queue. Learn the enterprise onboarding path, guardrails, BYOK and in-region routing hands-on. Build the APAC demo workspace. Agree the handoff format and overlap hours with US Support.
Lead discovery and PoCs on the APAC pipeline. Ship the residency data-path document. Turn the first month's tickets into macros and docs fixes.
PoC template with readout, account health review for every APAC account, first product feedback note with deal evidence, and a plan for the next regional hire.
Public sources only, read on 10 October 2026 (public job posting, 2026). Counts come from OpenRouter's public, keyless API. The demo is my own prototype built for this application, and its routing simulation follows the documented rules; the production router has more inputs.
API: models · providers · ZDR endpoints · EU models · US models
Routing: provider selection · in-region routing · ZDR · BYOK
Errors: errors and debugging · router metadata
Commercial: pricing · enterprise · docs FAQ · trust centre
Company: State of AI · Series B · Joining Stripe
Competitors: Vercel AI Gateway · Cloudflare AI Gateway · LiteLLM · Alibaba Model Studio · BytePlus ModelArk
Independent homework for the OpenRouter Customer Engineer APAC role · 2026 · edwardtay.com