The inference wars move to the silicon layer

DeepSeek hits gold-medal reasoning, Kimi k2.5 tops base model evals, and Nvidia answers Groq with a dedicated inference chip. The frontier is shifting from training scale to inference economics.

The model leaderboard got crowded this quarter, but the more interesting story isn't who's on top — it's where the competition has moved. Training-time scaling dominated 2024 and 2025. Now the real fight is over what happens when the model actually runs.

The reasoning tier keeps compressing

DeepSeek's V3.2 release claims GPT-5-level performance on its daily-driver variant and gold-medal results at IMO, CMO, and ICPC World with V3.2-Speciale, which the team positions against Gemini 3.0 Pro. That's a Chinese open-weights lab credibly claiming parity with the two most expensive closed frontier models. The underlying architecture matters as much as the scores: V3.2 debuted DeepSeek Sparse Attention, which let them cut API prices by more than 50% while extending long-context efficiency.

Meanwhile Perplexity's Aman Sanger reports that Kimi k2.5 was the strongest base model they evaluated, strong enough to justify a 4x scale-up in continued pre-training and RL. Two years ago the base-model conversation was OpenAI, Anthropic, and Google. Today the most interesting base models are coming out of Beijing and Hangzhou.

Silicon is catching up to the software

At GTC 2026, Nvidia announced the Groq 3 LPU, an inference-specific chip designed to pair with Rubin GPUs. The naming is pointed — Groq the company built its business on the argument that inference needs different silicon than training, and Nvidia just validated the thesis by shipping a chip with Groq's name on it. At the edge, Adtran's Ensemble Cloudlet is selling multi-node GPU infrastructure for on-prem inference, and researchers are squeezing KV caches with vector quantization to push more tokens through less memory. xAI's Grok 4.20 landed on Vertex AI marketed explicitly on inference speed for agentic tool calling, not reasoning depth.

What it means

For builders, the practical upshot is that model choice is no longer a binary between frontier-closed and good-enough-open. DeepSeek and Kimi are frontier. For buyers, inference cost is about to fall faster than model quality improves, which changes the ROI math on agentic workflows that were borderline last year. For creators and product teams, the latency floor is dropping — real-time multimodal experiences that felt aspirational in 2025 are becoming table stakes.

The bottom line: Stop benchmarking models on quality alone — the 2026 decision is quality-per-dollar-per-millisecond, and the leaderboard looks completely different through that lens.

Sources