Nvidia’s funding seen as key for US AI firms to compete with China

Building a frontier AI model in America is expensive. Training a model like Reflection AI’s Beam is estimated to cost around $100 million, a price tag that may make it harder for US labs to close the gap with China.

That is where Nvidia comes in. The chipmaker has turned itself into something like the sector’s venture fund, landlord and hardware store at once.

A $100 million model and a $6 billion bet

Reflection AI unveiled Beam on October 5, 2026. It is a sparse model with 501 billion parameters, of which only 23 billion are active at any given moment.

Beam was trained on 23.8 trillion tokens using reinforcement learning techniques. The run used 10,500 Nvidia GB300 GPUs over four weeks.

Beam is touted as delivering 3-4x better inference efficiency than China’s GLM 5.2. It also claims more than 4x the efficiency of leading Western open models.

Inference is the work a model does after training, when it actually answers prompts. Better inference efficiency means each answer costs less to serve.

The model is positioned as a cost-effective Western alternative, though it still shows some raw performance gaps against certain Chinese counterparts.

In August 2026, Nvidia struck a $6 billion licensing and investment deal with Poolside. The goal is to establish a robust ecosystem for open-weight models in the US.

Why China has the cost edge

The core problem is a pricing mismatch. Many Chinese models reportedly cost in the low millions to train. Leading US models like Beam are estimated at around $100 million.

Chinese labs have been shipping highly capable, low-cost models. Families like DeepSeek, GLM and Kimi have put real pressure on Western developers.

Nvidia as the sector’s financier

Nvidia’s strategy goes well beyond selling chips. The company has been taking equity stakes in emerging AI firms. Its total equity holdings in AI startups reached approximately $99 billion by mid-2026.

It is also building compute financing channels aimed at mobilizing upwards of $500 billion, and widening access to scarce GPU resources for the startups it backs.

What this means for the AI race

A large share of US open-weight ambition now runs through one company’s capital and chips. Nvidia’s backing may soften the blow of high training costs for startups, but a $100 million model still has to earn back $100 million against rivals that reportedly spent a fraction of that.

That is why Beam’s efficiency claims are the number to watch. If the projected 3-4x inference gains over GLM 5.2 hold up in real deployments, US labs gain a credible argument to compete on total cost of ownership rather than training cost alone.

Nvidia’s $6 billion Poolside deal and its approximately $99 billion startup portfolio give US labs breathing room. Whether that room turns into a self-sustaining ecosystem, or remains propped up by one patron, will shape how competitive American open-weight AI looks in the years ahead.