These startups are chasing the next big thing in LLMs

The article argues that transformers—the core technology behind today’s large language models—are increasingly a bottleneck on efficiency and scalability. It highlights the $50B computing spend projected by OpenAI this year and cites the International Energy Agency forecast that data-center electricity use will double by 2030, framing transformer-related power costs as a key constraint. Multiple startups are pursuing alternatives (sparse attention, context “power retention,” liquid neural networks, diffusion-based text generation, and state-space models), with early performance claims ranging from “rivals” mainstream LLMs on select tasks to benchmarks where Pathway’s Dragon Hatchling solved 97%+ of very hard sudoku puzzles.

Analysis

The equity read-through is less about “new AI winners” and more about who owns distribution if model economics reset. If cheaper architectures work in production, the value migrates from raw parameter scale toward product surfaces, cloud attach, and proprietary data loops — that’s constructive for GOOGL and, to a lesser extent, BABA, because both can absorb architectural change without needing the market to believe one single model family is the moat.

The bigger second-order issue is margin structure. Lower inference cost should widen gross margin on AI features, but it also risks commoditizing standalone model vendors and weakening the “compute scarcity” narrative that has supported AI infrastructure multiples. Near term, this is mostly a research headline; over 1-3 months, the market will care only if benchmark claims translate into lower serving cost, faster latency, or visible product adoption in search, cloud, or commerce. Over 6-18 months, a credible non-transformer stack could compress the premium on pure-play AI software while reinforcing integrated platforms.

Contrarian view: the consensus may be overestimating how binary this is. Even if transformers are no longer the dominant architecture, incumbents can borrow the gains quickly, so the trade is not “old AI loses/new AI wins” but “distribution wins, abstraction layer loses.” The main falsifier is real-world failure: if the new architectures don’t hold up outside narrow tasks, capex plans and AI monetization assumptions stay intact and the market should fade the headline impact.

Key Decisions for Investors

• Long GOOGL on weakness over a 1-3 month horizon; thesis is that cheaper AI serving improves search/ads economics and strengthens Cloud margin. Falsify if Vertex/AI attach metrics or cloud growth decelerate despite continued model rollouts.
• Add a smaller-sized long BABA as an optionality trade on efficient open-source model adoption in a cost-sensitive ecosystem. Keep sizing modest given geopolitical noise; stop if core commerce/Cloud trends do not show any AI-driven monetization benefit over the next 2 quarters.