Downlink

What do they actually do

Downlink provides a drop-in API that sits in front of existing, OpenAI-compatible LLM clients. Developers point their client’s base URL to Downlink and use a Downlink API key; the service then handles request routing and optimization behind the scenes while returning chat/completions through a compatible interface source.

Today, access is invite-only via a request form, and the company is at an early stage (YC W24; listed as a one-person team). The website claims it can boost performance (e.g., 20%+), expand rate limits, lower response times and costs, and select/tune models for customer use cases, but detailed mechanisms and independent benchmarks are not published yet sources: homepage, YC listing.

Who are their target customer(s)

• Early-stage product teams building LLM-powered features: Slow responses and low rate limits degrade user experience and block rollout of LLM features; they want faster, more reliable calls without heavy integration work source.
• Mid-size engineering teams with growing LLM usage: High, unpredictable API spend and throughput limits make it hard to scale users without costs rising linearly; they need better performance per dollar source.
• ML/platform engineers responsible for model ops: Time-consuming manual model selection, fine-tuning, and operational maintenance distract from shipping product; they want automation and a single API source.
• Teams already using OpenAI-compatible clients: They want to experiment with providers and optimizations but avoid rewriting integrations; a drop-in proxy minimizes migration risk source.
• Domain-focused teams (legal, healthcare, finance): Generic models underperform on specialized tasks; they need easy, managed fine-tuning for better accuracy without building infra in-house source.

How would they acquire their first 10, 50, and 100 customers

• First 10: Founder-led pilots with YC and personal network contacts; hands-on onboarding and custom tuning to prove measurable performance gains, using free credits and short pilot agreements to collect metrics and quotes homepage, YC.
• First 50: Broaden to an invite program with public examples, migration guides, and case studies; target developer communities and ML/platform engineers and introduce small referral incentives homepage.
• First 100: Open self-serve with pricing, playground, and full docs to enable bottom-up adoption; add partnerships and targeted sales for mid-market teams that need SLAs and managed fine-tuning, leveraging published benchmarks to shorten cycles homepage.

What is the rough total addressable market

Top-down context:

Downlink participates in the LLM infrastructure and optimization layer—routing, inference performance, and managed tuning—for teams building LLM features. The near-term addressable spend is a slice of what companies already pay for LLM APIs and supporting infrastructure.

Bottom-up calculation:

If 3,000–10,000 teams adopt an LLM gateway/optimization layer over the next few years with an average contract of $15k–$60k ARR (depending on usage and SLAs), the TAM would be roughly $45M–$600M. A mid-case (5,000 teams at $30k ARR) implies ~$150M.

Assumptions:

• Adoption measured as teams with material LLM usage that benefit from routing/optimization
• Average ARR varies by size/throughput and support level (self-serve vs. SLA)
• Downlink captures a portion of LLM infra budgets rather than full model spend

Who are some of their notable competitors

• OpenRouter: Router and unified API across many LLM providers; makes it easy to switch models and providers with one endpoint—overlaps with Downlink’s drop-in routing positioning site.
• Portkey: LLM gateway with routing, caching, observability, and failover across providers; focuses on reliability and cost/performance controls site.
• OpenPipe: Helps teams fine-tune and distill models to reduce latency and cost while maintaining quality; adjacent to Downlink’s “managed fine-tuning” value prop site.
• Cloudflare AI Gateway: A gateway for AI traffic that provides caching, rate limiting, analytics, and observability—useful for performance and cost control at the edge docs.
• Fireworks.ai: LLM inference platform focused on high-throughput, low-latency serving of open and proprietary models; competes on performance and cost for production workloads site.