Making fak attractive to India- and China-based startups (2026-06-30)

Concept note / go-to-market strategy. Answers the /goal: how do we make fak attractive to Indian-based startups — a wide range of things, including Hindi doc entry points, and the same for China. It maps the full lever set, grades each as shipped here / partial / proposed, and names the honest gaps. The first lever — in-language entry points — ships alongside this note under docs/i18n/ as the concrete proof, not a promise.

Why these two markets, and why fak fits

India and China are the two largest pools of startup engineering talent outside the US, and both build agent products under constraints that happen to line up with fak’s core design — not with a hosted US SaaS:

  1. Cost is denominated in USD, revenue is not. A token bill that is a rounding error for a US Series-B is a real margin line for an INR- or CNY-revenue seed startup. fak’s whole value proposition is the same agent loop, cheaper — cache reuse across a fleet (~4.1× less work than a tuned warm-cache stack on the 50×5 run, ~60× vs. the naive re-send loop), fewer wasted turns, and per-aspect routing that sends the cheap aspects to a cheap model. Margin, not features, is the pitch.
  2. Data must often stay in-country. India’s DPDP Act (2023) and China’s PIPL / DSL / CSL push personal and “important” data toward local processing and constrain cross-border transfer. fak is self-host-first: one static binary in front of a local model or a domestic provider, with fail-closed residency across backends, a default-deny capability floor, and a tamper-evident decision log. That is a compliance story a hosted API cannot tell.
  3. The models they already reach for are the ones fak already governs. China’s default open models — Qwen (Alibaba), GLM (Zhipu), DeepSeek, Yi, Baichuan, Kimi — are exactly the family fak fronts, and Qwen2/Qwen3 + GLM-MoE are proven bit-exact in the in-kernel reference engine (the rest are fronted over the OpenAI-compatible wire). fak is not asking a Chinese startup to switch models; it wraps the ones they run.
  4. Adoption has no payment rail to cross. fak is Apache-2.0, free, self-host. There is no cross-border card-billing / entity / invoicing friction — the quiet tax that slows hosted-SaaS adoption in both markets. git clone and go install are the whole funnel.

The lever set (wide range), graded honestly

# Lever India China Status
1 In-language doc entry points Hindi Simplified Chinese Shipped here — docs/i18n/ POC pages
2 In-language answer-engine (AEO) discovery Hindi + English Baidu/Zhihu/Juejin, zh terms Proposed — extends internal/marketing/aeo.go
3 Domestic-model fit open + IndiaAI models Qwen/GLM/DeepSeek/Yi/Kimi Partial — Qwen/GLM in-kernel; rest over the wire
4 Data-residency / compliance framing DPDP Act 2023 PIPL / DSL / CSL Partial — capability exists; framing is new
5 Cost framing in local unit-economics INR margin CNY margin Proposed — re-skin the benchmark story
6 Domestic silicon via the compute HAL nascent (IndiaAI compute) Ascend / Cambricon / Biren / Moore Threads Partial — HAL ships CUDA/Vulkan; a domestic backend is a registration seam, not a shipped backend
7 Access / distribution friction low (GitHub OK) GOPROXY, ModelScope, Gitee mirror Partial — GOPROXY tip is real & shipped in zh page; mirrors proposed
8 Zero payment friction (Apache-2.0, self-host) ✓ ✓ Already true — no work needed, just say it
9 Community channels dev communities, IndiaAI, campuses WeChat, Zhihu, Bilibili, Juejin, Gitee Proposed — GTM action, not code
10 Per-market positioning line “cheaper agents on your stack, data in-country” “govern + cache the domestic models you already run” Proposed — copy, lands in the i18n pages

1 — In-language documentation entry points (shipped here)

Native-language entry points do two things at once: they lower the first-contact barrier for a non-native-English reader, and they signal welcome — a project that put a Hindi and a 中文 front door up is visibly courting these developers. The honest scope is an entry point, not a full doc-set translation: a compact, faithful page that carries the one-line pitch, the 60-second proof, the install path, and the market-specific value props, then hands off to the (English) deep docs. This is what ships with this note:

Fence: these are machine-authored translations pending native review — the hub marks them as such and asks for corrections via issue/PR. Do not claim “professionally localized.” Update 2026-07-01: the Indian-language follow-ons shipped — Tamil (ta/), Telugu (te/), Bengali (bn/), Marathi (mr/) — each with the same machine-authored fence. Remaining follow-ons: native review of all pages and Traditional Chinese (TW/HK).

2 — In-language answer-engine discoverability

The repo already runs an AEO program (internal/marketing/aeo.go, tools/seo_aeo_scorecard.py, gen_structured_data.py) so LLM answer engines surface fak for the right English terms. Chinese developers ask Baidu / Zhihu / 掘金 (Juejin); many Indian developers search in Hindi-English code-switch. The lever is to emit in-language disambiguation terms and structured data so an answer engine responding in Hindi or Chinese names fak for “agent 内核 / 工具调用防火墙” or “एजेंट कर्नेल”. Gap: no localized terms are emitted today; this is a bounded extension of the existing generator, not new infrastructure.

3 — Domestic-model fit (fak’s strongest China hook)

China’s agent startups overwhelmingly build on domestic open models under a “self-host-first” posture. fak’s position is not “use our model” — it is “keep your model; we govern and cache it.” Proven bit-exact in the in-kernel reference engine: Qwen2/Qwen3 and GLM-MoE. Fronted unchanged over the OpenAI-compatible wire (Ollama / vLLM / SGLang / llama.cpp / LM Studio, or a domestic API): DeepSeek, Yi, Baichuan, Kimi, and any open-weights model. For India, the same wire fronts open models and the emerging IndiaAI / Sarvam / Krutrim endpoints. Honest split: in-kernel decode is proven only for the listed architectures; everything else gets the full capability / cache / audit / route value over the wire, just not in-kernel decode.

4 — Data-residency & compliance framing

fak already has the mechanism (self-host, fail-closed residency, default-deny floor, X-Trace-Id-correlated audit log). What is missing is the framing aimed at these two regulatory regimes:

This ships here as the data-residency & compliance explainer — the mechanism mapped to both regimes, with the honest “not legal advice, not a certification” fence. No new code; the boundary already had the properties.

5 — Cost framing in local unit-economics

Re-tell the existing (honest) benchmark story in the language a cost-sensitive founder reads: cost per 1,000 agent turns, margin per seat, the self-host fleet reuse win. Keep the net-true-value discipline — quote the tuned baseline (~4.1×), not the naive 60×, as the headline. No new measurements; a re-skin of BENCHMARK-AUTHORITY.md numbers.

6 — Domestic silicon via the compute HAL

Under export controls, Chinese startups increasingly target Huawei Ascend, Cambricon, Biren, Moore Threads; India’s domestic compute (IndiaAI) is nascent and mostly NVIDIA today. fak’s internal/compute HAL + neo-silicon onboarding is the seam to add a compute.Backend by registration rather than re-forking the forward pass — a genuine “bring-your-accelerator” story. Fence: the HAL ships CUDA and Vulkan backends; a domestic-accelerator backend is a registration path we support, not a backend we have shipped or benchmarked. Say “the seam is ready,” never “we run on Ascend.”

7 — Access & distribution friction (China-specific, partly real today)

8 — Zero payment friction (already true — just say it)

Apache-2.0 + self-host means no card, no cross-border invoice, no entity stands between a developer and adoption. This removes the single biggest hosted-SaaS adoption tax in both markets. It costs nothing to ship because it is already the license; the lever is to state it prominently in the localized pages (both do).

9 & 10 — Community channels and per-market positioning

Go-to-market actions, named for completeness, not built here: seed the China-facing channels (WeChat 公众号, Zhihu, Bilibili, 掘金, a Gitee mirror) and the India-facing ones (dev communities, IndiaAI, campus/hackathon ecosystems), each with the per-market positioning line from the table. These belong to a human GTM owner; this note’s job is to make the docs and the product ready for them.

What ships with this note vs. what is the next checkable step

Ships now (docs-only, no code, green under the doc gates):

Not yet — the honest follow-ons, each a bounded next step:

Reported as not yet where unproven, per the repo’s honesty contract: the mechanism for every lever above already exists in fak; what is new here is the framing, the in-language front doors, and one accurate onboarding fix — not a claim that any market has adopted it.