alex atala

Token price reductions trigger super-linear usage growth; OpenAI's 10x price cut on Luna via OpenRouter yielded 13x usage increase, demonstrating inference demand elasticity.

Chinese open-source labs (Moonshot, Zhipu, DeepSeek) benefit from national champion dynamics, unlimited funding, regulatory freedom, and top-tier researchers that Americans underestimate; US open-source labs face harder fundraising and competition from frontier labs.

Meta's Muse can leverage its social graph and people-focused brand to differentiate from pure model labs; they have the compute and talent to compete once they find their niche.

Zhipu's GLM 5.2 was a significant step forward for open-weight model quality, narrowing gap with frontier models; Moonshot's Kimmy K3 is catching up to that level.

Inference providers doing custom hardware and low-level optimizations (like Fireworks) are structurally advantaged; they also lead on making model customization easier via portable adapters.

DeepSeek has become China's designated AI champion, concentrating state resources and regulatory freedom; they and ByteDance are aggressively pursuing custom chips to overcome export controls.

OpenRouter's 100% focus on routing/gateway/marketplace creates durable advantage over companies treating it as a side feature; enterprise pricing with committed spend aligns incentives as inference demand grows 10-15x annually.

Nvidia's top priority is preventing hyperscaler GPU concentration; they actively want heterogeneous inference provider competition which creates durable demand for their chips across many customers.

Poolside is a promising new American model lab focused on small but highly effective coding models with useful developer tooling.

Anthropic launched Claude Design not for direct revenue but to make design teams dependent on their models, creating a strategic foothold inside enterprises that competes with app-layer companies like Figma.