What leaders & investors need to know about the changing economics of AI compute
In the US, the AI titans (OpenAI, Anthropic, Meta, etc) are threatening to disrupt Fortune 1000 incumbents via offerings like Cowork and Altruist. But in turn - new low cost models from China (and there will be others no doubt) are threatening the AI titans.
At the same time, AI is turning into a political issue for the US midterm elections. But this is more than US-China geopolitics as usual. Three competing AI architecture patterns are emerging. Leaders and investors alike need to understand their differences, and the implications for ROI.
First, the latest low cost models from DeepSeek's competitors...
This week (of February 15) saw the rollout of a number of low cost high performing models from China. Last year around this time, Deep Seek shocked the world with its models showing that you don't have to be a Silicon Valley giant to turn out a respectable model. This year the rest of the competitive field are showing what they can do.
Four are particularly notable:
Moonshot AI — Kimi-1.6 / Kimi-Reasoner
• Official model portal: https://kimi.moonshot.cn
• Model listings & weights (Hugging Face): https://huggingface.co/moonshotai
Moonshot expanded its Kimi family into reasoning-tuned models optimized for long-context and coding agents, emphasizing cost-efficient inference and enterprise deployment rather than leaderboard performance. The company is explicitly targeting real-world task completion per dollar following the efficiency model demonstrated by DeepSeek.
🔥Kimi also just released a one-click cloud deployment for OpenClaw.
Zhipu AI — GLM-4-Open
• Model page: https://open.bigmodel.cn
• Hugging Face release: https://huggingface.co/THUDM
• Academic background (Tsinghua GLM research): https://arxiv.org/abs/2210.02414
Zhipu released an open-weights version of its GLM-4 reasoning model designed for private datacenter and sovereign deployments, allowing organizations to run a GPT-class system without reliance on U.S. cloud APIs. The project comes from the Tsinghua University GLM research lineage, one of China’s most mature alternative foundation-model ecosystems.
01.AI — Yi-Large-MoE-Open
• Model repository: https://huggingface.co/01-ai
• Company announcement: https://01.ai
01.AI (Kai-Fu Lee’s lab) released mixture-of-experts Yi variants aimed at enterprise self-hosting, positioning them as a multilingual, fine-tunable alternative to Meta’s Llama models. The notable feature is enterprise-friendly licensing plus efficient inference, making the models practical for regulated industries.
Alibaba Cloud — Qwen-MoE-Open expansion
• Official Qwen documentation: https://qwenlm.github.io
• Model weights: https://huggingface.co/Qwen
• Technical paper: https://arxiv.org/abs/2309.16609
Alibaba expanded its open Qwen family with mixture-of-experts variants and smaller deployable builds, focusing on on-prem and hybrid-cloud usage rather than pure cloud APIs. The strategic angle is providing a complete enterprise AI stack that customers can operate at predictable cost.
These releases all follow the same playbook: open or open-weights reasoning models + mixture-of-experts efficiency + local deployment.
The economic assumptions of US AI tech giants now under scrutiny
In the US, the AI leaders (Anthropic, Meta, OpenAI, Microsoft) have largely followed a conventional wisdom about ROI model:
- Training larger models on enormous hyperscaler training clusters will create durable competitive advantage, and barriers to entry for other players.
- Large centralized platforms (enterprise focused or social focused) will host millions of users who come into walled gardens and use services there.
- Platform owners extract fees from users based on consumption
- Consumption based fees enable a return on investment for the exorbitant data center buildouts and energy consumption.
All four points of this thesis may be at risk of disruption: If low cost, high-performing models can be trained without hyper scale infrastructure and best-in-class chips, and if they can run on one's own hardware, then why would anyone continue to pay token fees to a large centralized player?
Kristin O’Donoghue notes here that the number of U.S. professionals adopting the title of “founder” on LinkedIn increased 69% in 2025 and is up 300% since 2022. The average age of entrepreneurs is dropping as AI speeds up key aspects of the founder's role.
How many of these are paying for subscriptions to OpenAI or Anthropic vs. running open weights LLMs on their own hardware?
In my own experimentation, I'm already beginning to move towards running models on my own hardware in order to avoid fees and control spend, and there is anecdotal evidence that other organizations are leaning in the same direction.
So is this just the geopolitical AI arms race playing out?
In past editions of S3T we've noted AI a geopolitics arms race to secure AI dominance (for military superiority and economic flourishing) and to secure the necessary energy, materials, technology and talent to win, or at least be among the winners. You certainly could view these recent model releases as China's AI ecosystem following a geopolitical playbook - motivated in part by necessity (tech sanctions) and opportunism - that undermines US AI companies.
That may be accurate on some level, but there's actually something more going on...
3 AI architecture patterns are emerging:
OpenAI, Anthropic, and Chinese AI companies are actually driving three competing philosophies of AI compute - and notice how they differ in terms of where intelligence lives and where data lives: