AI Competition between the US and China is in Open Source models

The Financial Times reports (Oct7, 2026) that a US AI start-up backed by Nvidia and the Trump admin­is­tra­tion has unveiled its first open-weight model, boost­ing Amer­ican ambi­tions to com­pete with Chinese groups that dom­in­ate the mar­ket for pub­lic, cus­tom­is­able AI products.

Reflec­tion AI claims that its first model, Beam, is on a par with Z.ai’s GLM-5.2, a lead­ing Chinese com­pet­itor.

My view is that without political sanctions, its not a market in which the US has the advantage given China's headstart. So called open weight models provide a method for customisation at lower cost for tech companies prepared to build their own AI applications rather than relying on expensive general models provided by the big AI companies. Nvidia wants to see this market segment grow to bolster demand for its chips with the objective of monopolising this market.

The start-up is stra­tegic­ally sig­ni­fic­ant for the US in its AI race with China as well as for Nvidia, which is aim­ing to cre­ate rivals to the biggest labs includ­ing OpenAI and Anthropic.

Reflec­tion said Beam out­per­formed US peers, “deliv­er­ing com­pet­it­ive per­form­ance on cod­ing and agen­tic tasks at a frac­tion of the token cost”, but had not yet pub­lished full tech­nical details of the model’s per­form­ance. Tokens are units of inform­a­tion pro­cessed.

Reflec­tion has partnered with the Pentagon and the US Depart­ment of Energy as well as strik­ing agree­ments to develop AI mod­els for US allies includ­ing South Korea as part of the Trump admin­is­tra­tion’s efforts to fend off Chinese com­pet­i­tion and expand its own sphere of influ­ence.

Reflec­tion raised more than $4bn since its launch in 2024, racing to a $25bn valu­ation before releas­ing its first model.

Open-weight mod­els, which users can cus­tom­ise and run on their own hard­ware, can be far cheaper for busi­nesses to oper­ate than the most advanced ver­sions of OpenAI’s Chat­GPT or Anthropic’s Claude.

The search for a lead­ing US open­weight model has become more urgent in recent months as usage of the tools has surged with Chinese labs behind most top mod­els. In August, open­weight mod­els accoun­ted for 56 per cent of all tokens pro­cessed through AI cloud applic­a­tion com­pany Ver­cel’s AI Gate­way, up from 7 per cent in Decem­ber.

Some Trump admin­is­tra­tion offi­cials — includ­ing Treas­ury sec­ret­ary Scott Bes­sent — have floated the idea of impos­ing sanc­tions on Chinese mod­els thought to be steal­ing IP from Amer­ican com­pet­it­ors.

The threat promp­ted an industry­wide let­ter, led by Nvidia, which stated that an open-weight “eco­sys­tem” was essen­tial for “cre­at­ing oppor­tun­it­ies for innov­a­tion and prosper­ity across the coun­try”.

Nvidia boss Jensen Huang has also defen­ded the use of China’s open-weight mod­els by Amer­ican busi­nesses, say­ing they provide good value for com­pan­ies.

Nvidia has inves­ted $800mn into Reflec­tion and provided access to its chips, help­ing the com­pany gain ground on Chinese rivals. Reflec­tion has also been backed by ven­ture firms includ­ing Sequoia Cap­ital, Dis­rupt­ive and 1789 Cap­ital, at which Don­ald Trump Jr, a son of Pres­id­ent Don­ald Trump, is a part­ner.