Get ready for US open LLMs and just in time

The teeth gnashing and near panic over open source large language models from China has reached a fever pitch as Z.ai, DeepSeek, Alibaba's Qwen and Moonshot AI's Kimi close the gap on US proprietary frontier models and squeeze projected (fabricated to some) future profit margins of AI giants.

While advances in Chinese AI have executives from Anthropic and OpenAI spooked and pushing ways to restrict open source models, the reality is that US versions of open LLMs are likely to surge too. Enterprises will be pleased either way since you can't have a model duopoly. Just because OpenAI and Anthropic have $1 trillion private valuations doesn't mean you have to pay for it.

In recent days, we've seen the following:

• Poolside AI launched Laguna S2.1, which is designed for efficiency and longer horizon work. Laguna S2.1 is a 118B total parameter mixture-of-experts model with 8B active parameters per token. The new model appears to hold its own against the popular open weight models.
• Thinking Machines launched Inkling, which is the first model in what'll be a family of LLMs that are open and set up to be customized.
• Nvidia's Nemotron family continues to perform well on the popularity charts headlined by Nvidia Nemotron 3 Ultra. Palantir partnered with Nvidia to lower costs. See: Nvidia Nemotron: Much needed open-source model champion in US | Nvidia launches Nemotron 3 open models to enable multi-agent systems
• Toss in open models from Mistral, Google and OpenAI, which has a few, and there's a base to work with if you wanted to avoid Chinese AI models. The reality though is the open source innovation in from the China labs due to compute constraints. China labs simply had to be more creative. '=
• Cisco launched Antares, a family of security small language models built for pinpointing where vulnerabilities exist in a code base. The models Antares-350M and Antares-1B were released on Hugging Face. The big takeaway is that Cisco claims Antares is 172x cheaper than OpenAI's GPT-5.5 and 15.2x cheaper than Z.ai's GLM-5.2. Smaller models are a threat to the larger pricey versions that overthink every task.

Given the joke that tokenomics is and the reality AI bills are too high and the general vibe that no one in the AI ecosystem wants two players to drive the frontier model development it's not surprising that competition is going to ramp. SpaceXAI and Meta already see the pricing from OpenAI and Anthropic as opportunity. See: SpaceXAI, Meta puts pricing squeeze on Anthropic, OpenAI | Where’s tokenomics for the rest of us?

Amazon Web Services and Google Cloud also see the need for commodity models. Salesforce is also a proponent of smaller models and tailoring them for tasks. Microsoft is also pushing the commodity model play as it has been whacked by its own token bills. Rightsizing open models may cut your AI inference spend

Those three hyperscale cloud providers could easily squeeze down costs and punch Anthropic and OpenAI. What's the problem? All three are investors in those two LLM leaders. Some free advice: Anthropic and OpenAI should both launch open weight models to upsell later.

In second half of 2026, my bet is we're going to see a surge in US-based open model providers. You can fret over China AI labs today and even limit them, but the problem for profit margins Anthropic and OpenAI is only beginning. The timing of this competition is unfortunate for the IPOs of those two companies, but that's not your problem.

• Moonshot AI launches Kimi K3
• Is a FINRA-ish group the answer for frontier AI standards?
• DeepSeek's real impact happening now
• Here’s what we learned about AI projects from enterprise buyers so far
• AI inference costs are going to be a big concern: What's the fix?
• Why enterprise AI leaders need to bank on open-source LLMs