OneTriangle

Independent validation is thin. The only third-party surface found is an auto-generated directory, LLM Relevance, which groups OneTriangle with near-peer serving platforms such as Baseten, Runpod, Tensorlake, Gobii, Concentrate, and AgentZ; no analyst or benchmark comparison against vLLM, Fireworks, or Together was located, so headline gains rest largely on vendor-reported numbers.

Structurally the company is two founders plus roughly four engineers, entered inference in mid-2026 after pivoting from the TrustAI agent-compliance product, and lists no named enterprise customers, SLAs, or certifications. Its own pricing page concedes that Llama 70B rates are GPU-only breakeven at fully productive throughput, excluding idle capacity and overhead, which narrows the margin story versus scaled hyperscalers.

External commentary points the same way as the company's own framing: Introl's infrastructure analysis projects inference at 65% of AI compute by 2029 and 80-90% of lifetime AI costs, while agent workloads generate on the order of ten times the tokens of a chat session. Dataintelo characterizes the inference optimization platform market as intensely contested among large vendors, specialists, and startups, confirming both demand growth and competitive density.

OneTriangle's stated bet is architectural: model-mixing becomes the default serving pattern, with small models handling prefill and routine tokens, large models handling hard tokens, and cache moving freely between them. If that pattern standardizes, the transfer layer captures inference economics; if incumbent serving platforms absorb the technique, the advantage compresses to execution speed.

The thesis targets the largest line item in AI compute: the founders assert inference has surpassed training as AI's biggest compute cost, absorbing up to 90% of AI budgets, with prefill dominating on long-context and agentic workloads. The claimed wedge is proprietary cross-model KV cache transfer delivering about 20% lower prefill cost and 40% faster TTFT, with internal figures suggesting parity with or better divergence than NVIDIA's published approach, achieved within weeks.

Bull points are the vLLM upstreaming path and advice from that project's lead, plus a deep MIT research bench. Bear points are that evidence is mostly self-reported, the market already hosts well-funded serving platforms, and the company pivoted from its TrustAI compliance product in 2026 with roughly four staff and no disclosed customers; diligence should demand reproduced benchmarks and reference buyers before weighting the headline claims.

OneTriangle prices per million tokens on six hosted models, with input rates from $0.021 (Qwen3 4B Fast and the 8B safety classifier) to $0.15 (DeepSeek V4 Flash) and output rates from $0.35 to $6.43 for the Llama 3.3 70B Fast endpoint. Cached input carries its own discounted rate on most models, so reused context is cheaper than fresh prefill.

Two levers shape the strategy: a Flex/Delayed tier priced 40% below Priority for delivery within ten hours, which monetizes spare GPU capacity, and beta-free vision and robotics models on dedicated GH200 runtimes to seed new workloads. The page notes Llama 70B rates sit at GPU-only breakeven, signalling deliberate penetration pricing on flagship open-weight serving.

OneTriangle operates a KV-cache-aware inference engine that splits long-context serving into two stages: a small model handles prefill, then its key/value cache is projected into a larger model's attention space so decode starts against context it never computed. The company reports roughly 20% lower prefill cost and about 40% faster time-to-first-token on validated model pairs such as Llama 3.1 8B-to-70B and Qwen3 8B-to-32B.

Buyers reach it through an OpenAI-compatible API and can also buy managed hosting of open-weight families including Llama, Qwen, Mistral, Gemma, and DeepSeek, with the provider running GPUs and serving while the customer changes one line of configuration. Six production endpoints plus beta vision and robotics models are listed, priced per million tokens.