Good Enough Wins: Open-Weight Models Now Carry Most of AI's Workload

The Moment Everything Changed

For three years the AI business ran on one assumption: the best model wins, and customers pay a premium for it. This week the data said otherwise. Vercel’s latest production index found that open-weight models handled 71.5% of the tokens on its AI Gateway in September, and IBM’s chief executive told investors that models nearly as good as the frontier, at a fraction of the price, are now the rational choice for most enterprise work. Taken together, the week’s announcements show the model layer starting to look like a commodity, and the AI industry’s economics are being rewritten around that.

Background

Open-weight models, whose parameters anyone can download, run and modify, used to be a hobbyist’s compromise. They trailed the frontier by a wide margin, and serious production work went to closed APIs from Anthropic, OpenAI and Google. That gap has been closing for about a year. According to IT Brief’s coverage of Vercel’s September index, open-weight models carried just 7% of gateway tokens in December 2025 and 13% in April 2026, then jumped to 56% in August, the first month they held a majority.

Enterprise surveys tell a similar story. An Enterprise Technology Research survey of 200 organisations, reported by Techstrong.ai in September, found that open-weight models account for 34% of enterprise AI tokens, up from 23% a year earlier. Respondents expect that share to reach 41% within twelve months. Cost savings were the main reason given, cited by 69% of production users. The change is happening at different speeds in different places, but it is happening everywhere.

What Happened

The week opened with The Register’s report on Vercel’s October index. Open-weight models now carry 71.5% of gateway tokens, and spending is following. Their share of dollars rose from about one in twenty in June to nearly one in four in September, while the average price per token fell about 47%. Almost one in three teams changed their most-used AI provider in a single month. Anthropic’s Claude Opus 5.5 still took the largest share of September spending at 11.03%, but DeepSeek carried 58% of all tokens for just 6.6% of spend.

On 19 and 20 October, Amazon and Google went after the software layer that keeps developers on Nvidia hardware. Amazon released TorchNeuron for its Trainium chips and submitted code to decouple PyTorch features from Nvidia GPUs. Google open-sourced TorchTPU the next day, saying existing PyTorch apps can move to its Tensor Processing Units with a configuration change rather than a rewrite. As TechCrunch noted, that does not mean developers will abandon Nvidia en masse, but it lowers the cost of leaving.

Nvidia answered with openness of its own. At GTC Berlin it released a batch of Nemotron 3.5 models and, as The Register reported, published the training data and recipes alongside the weights. That lets European buyers wary of US-controlled AI audit the models and build fully custom variants. It also announced Nemotron 3.5 Bolt, a 74-billion-parameter mixture-of-experts model that fits in 24 GB of video memory, putting a capable model within reach of a single workstation card.

Later in the week IBM’s Arvind Krishna made the economic case on the company’s earnings call. In The Register’s account, he said open-weight models now deliver about 90% of frontier quality at 10 to 25% of the token cost, and that a useful deployment fits in two or three racks.

Why It Matters

When the second-best product costs a fraction as much and is nearly as good, pricing power erodes quickly. Most enterprise AI work is summarising, classifying, extracting and answering routine questions, and for that work the last ten points of benchmark quality rarely justify a tenfold price difference. Frontier labs keep their premium on the hardest tasks, such as coding agents and complex reasoning. But the bulk of the volume, which is what fills data centres, is moving to whatever runs cheapest.

The effects reach well beyond model vendors. Easy switching shifts bargaining power to buyers, which is why Replit could drop Anthropic as its default provider this month and reduce the decision to cost and benchmark results. Portable PyTorch weakens Nvidia’s software lock-in just as the company is reserving years of substrate and packaging capacity on the bet that demand will last. On-premises inference also brings back a market for traditional infrastructure vendors such as IBM and Red Hat, who had looked sidelined by the API era.

There is a geopolitical side too. Many of the cheapest high-volume models come from Chinese labs, and The Register found that more than two-thirds of September spending on Chinese models went to US providers serving them. In the same week, the Department of Homeland Security ordered federal civilian agencies to strip DeepSeek products from their systems. Commodity economics and national-security policy are now pulling in opposite directions.

Expert Perspectives

Krishna described how quickly the gap has closed. “In late 2025, we began to see that gap shrink, where maybe the gap was not 70 to 20 or 30, it was 70 to 50,” he said on IBM’s call, as quoted by The Register. He then put the buyer’s question bluntly: “Would you like something that’s 10 percent of the cost if it’s 90 percent as good as the best?” He also warned that the bottleneck is now people, asking, “Where are my kernel engineers?”

Chris McGuire, senior fellow for China and emerging technologies at the Council on Foreign Relations, flagged the supply risk. China has proposed adding 23 AI-related items to its export control list, and McGuire told The Register that “this introduces significant uncertainty for US businesses that host or rely on Chinese models.” Licensing terms set in Beijing could change the economics of a deployment overnight.

Not everyone thinks cheaper tokens automatically mean smaller bills. Writing in CIO, Microsoft senior product manager Shania Rasheed Nalagath argued that open weights are good news for buyers but that switching “still won’t lower your bill this year” without rigorous evaluation. “Demos flatter every model about equally,” she wrote, which is why informal tests can’t reveal what a cheap model is quietly costing in quality.

What to Watch

The first signal is frontier pricing. If Anthropic and OpenAI cut prices further or push cheaper tiers, they are conceding that volume is moving to commodity rates. The second is Nvidia’s margins. Its supplier deals assume demand for years, while TorchTPU, TorchNeuron and small models like Nemotron 3.5 Bolt widen customers’ options. The third is regulation of Chinese open-weight models, from export controls in Beijing to procurement bans in Washington, which could fragment the cheapest part of the market overnight.

Inside enterprises, the question is moving from “which model is best” to “what is good enough for this workload, and can we prove it.” Teams that build evaluation pipelines and the in-house skills Krishna says are missing will capture the savings. Those that don’t will just swap one invoice for another. The frontier still matters, but this week made clear that most AI work will now go to the cheapest model that is good enough.