You Don't Need the Smartest AI Anymore
Speakers
Josh Kale
— Co-Host, Limitless Podcast (Bankless)
Ejaaz Ahamadeen
— Co-Host, Limitless Podcast (Bankless)
Summary
The episode analyzes the shift in the AI industry from a single race for the smartest model to a two-market structure split between frontier intelligence and low-cost efficiency. The hosts examine new models from SpaceX AI (Grok 4.5) and Meta (MuseSpark 1.1), arguing that these companies are poised to capture enterprise demand with cheap, agentic models backed by massive compute investments and custom silicon. They also highlight implications for HBM memory demand and the emerging role of model routing.
- SpaceX AI's Grok 4.5 delivers frontier-class performance at a fraction of the cost, using parallel agents and a fully integrated stack.
- Meta's first closed-source model, MuseSpark 1.1, is the cheapest per unit of intelligence and leverages its own MTIA 400 silicon.
- Both companies possess large compute advantages and are aggressively building data centers, challenging incumbent AI labs.
- The AI market is diverging into low-cost workhorse models and premium frontier models for high-stakes tasks.
- Enterprise token expenditure is falling as companies switch to cheaper models, validating the demand for cost-efficient AI.
- Meta's custom chips significantly increase HBM content, pointing to higher demand for high-bandwidth memory.
- Model routing is expected to become a crucial layer, but no clear public company is positioned for it yet.
Ideas
SpaceX AI's cheap models capture enterprise.
SpaceX AI's Grok 4.5 is a frontier-class model that uses 4x fewer tokens than Claude Opus 4.8, resulting in 17x lower cost for the same task. The model employs parallel agents, leveraging the Cursor acquisition for intelligent request routing. SpaceX AI has a fully integrated stack (data centers, custom chips) and an aggressive pipeline of larger models (5–10 trillion parameters). This enables them to deliver near-frontier intelligence at a fraction of the price, positioning them to dominate the price-sensitive enterprise market.
Meta wins low-cost AI with custom silicon.
Meta's MuseSpark 1.1 is the cheapest model per unit of intelligence, costing 25% of frontier models and undercutting even the cheapest Chinese open-source model. It uses a master agent for orchestration, is omni-modal (video, images, text), and is fine-tuned for computer use. Meta is vertically integrated with its own MTIA 400 silicon, making massive data-center investments. This shift to closed-source and the company's compute advantage position Meta as a serious competitor in low-cost enterprise AI, potentially capturing the broad middle market.
HBM demand surges from AI chips.
Meta's new MTIA 400 chips use 51% more HBM, and the broader shift to custom AI silicon by hyperscalers will substantially increase demand for high-bandwidth memory. HBM is a direct beneficiary of the trend toward more memory-intensive AI accelerator designs.
