A GitHub project surfaced on Hacker News this morning showing DeepSeek V4 Flash 0731 running on a single AMD MI300X. The HN comment thread did what HN comment threads do: they ran the math.
The result is the clearest public quantification of API-vs-self-hosting cost economics that has appeared in months. And the conclusion — for most enterprises — is not what the self-hosting advocates claim.
From the project README and the HN thread:
One commenter pushed back: at $1.99/hour with mixed load, you might generate $3–$4/hour in tokens. Even in that optimistic scenario, the economics are close to break-even — not the 10× savings self-hosting advocates often claim.
The self-hosting math assumes 100% utilization. In practice, enterprise AI workloads are bursty:
At 30% utilization — a realistic number for a single-team deployment — the self-hosting cost per useful token triples: $1.99/hour ÷ 30% = $6.63/hour of actual output. The 3.6× gap becomes 10×.
API providers solve this through multiplexing: thousands of customers share the same GPU fleet, keeping utilization above 80%. The per-token price they charge reflects that efficiency. A single tenant cannot match it without dedicating engineering effort to utilization optimization that has nothing to do with their actual product.
On the same HN front page, another project appeared: Swiftlet, a Swift + Metal runtime running 80B Qwen in 4.3 GB RAM on a Mac and 35B on an iPhone 17. 214 points, 96 comments.
Swiftlet is technically impressive. It's also not a production routing tier. The decode speed — 4.5–5 tokens/second for the 80B model — is impractical for enterprise workloads where you need hundreds of tokens per second to serve a team. The iPhone result at ~1 token/second is a proof of concept, not a deployment strategy.
The HN thread itself captured the bifurcation: commenters acknowledged that on-device inference is progressing, but the consensus was that API providers have a permanent efficiency advantage from specialized infrastructure. The market is bifurcating:
Trimio's value proposition sits on the enterprise side of that bifurcation. The governance layer — virtual key attribution, budget enforcement, routing rules, audit logs — is what enterprises need when they're consuming API tokens at scale, not when they're running a model on a Mac.
The MI300X project confirms three things for enterprise buyers:
The HN thread provides a concrete, citable cost comparison that's more credible than any vendor whitepaper. The math is simple enough for a CFO to verify:
The Trimio layer adds: per-virtual-key cost attribution (which team consumed which tokens), budget enforcement (alerts when a team exceeds their allocation), quality-aware routing (V4 Flash for execution, Opus 5 for planning), and audit logs (every request traceable to a virtual key and request ID).
Self-hosting gives you: a GPU in a rack and a utilization problem.
The self-hosting debate has been running since the first LLM API launched. The MI300X project is the clearest data point yet: even with excellent hardware (192 GB HBM3, no quantization needed), impressive throughput (830 tok/s aggregate), and the cheapest per-GPU rental available ($1.99/hour), self-hosting is 3.6× more expensive than the API at saturation — and 7–12× more expensive at realistic utilization.
For enterprises evaluating self-hosting vs. API routing, the math is now public and verifiable. The conclusion is straightforward: API routing wins on cost, and the governance layer (attribution, budgets, routing rules, audit logs) is what makes API routing managed rather than just cheaper.
DeepSeek V4 Flash at $0.14/$0.28 per MTok is the cost floor. Trimio routes to it. Self-hosting doesn't beat it.
Trimio's LCR engine routes execution-tier workloads to DeepSeek V4 Flash at $0.14/$0.28 per MTok — with per-key cost attribution, budget enforcement, and audit logging included. See the routing table.