Latest: Open-Weight Models Hit 62% of Enterprise Tokens; Bain Flags $4.2T Revenue Gap
The growing gap between token volume and spending share shows enterprises routing high-volume, lower-stakes work to open-weight models at a fraction of frontier costs, while closed models retain the work where errors are expensive. The Bain revenue gap and Slok's distribution data raise concrete questions about whether the AI buildout can pay for itself across the broader economy.
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
Open-weight models reached roughly 62% of enterprise token volume on Vercel's AI Gateway by August 22, 2026, up from 11% in April and 36% in July, while capturing under 9% of spending, with closed models holding about two-thirds of dollars spent. Chinese-built models represent about 29% of tokens as a subset of that open-weight total. American companies including Figma and Pika are building products directly on Chinese open models such as Alibaba's Qwen and MiniMax. Bain's 7th Global Technology Report projects AI must generate $6 trillion annually by 2031 to cover infrastructure costs, with existing applications projected to produce only $1.2, 1.8 trillion, leaving a $4.2 trillion gap Bain attributes to products and uses that do not yet exist. Apollo economist Torsten Slok found revenue per worker and profit margins are broadly flat for companies outside the seven largest technology stocks, suggesting AI's economic gains have not spread widely.
Reports emerged that Figma post-trained its new model on Alibaba's Qwen, Pika built on MiniMax, and WeChat also built on Qwen, placing Chinese open models inside widely used American creative and social tools.
Seventeen new Chinese AI models shipped in a single week, including Kimi K3.1 Pro, with Qwen4 potentially arriving soon. Milk Road AI stated 'The United States still leads at the frontier, but China is winning the open model race'.
Mistral released weights for its 1-trillion-parameter Large 4 model in late October, roughly halving API prices but under a modified MIT license requiring MaaS providers to obtain a separate commercial license, and its Terminal-Bench 4.0 score lands one point above GLM-5.3.
Neobank bunq reported its Finn assistant runs on open-weight models via Nebius at 60% of closed-model costs while matching quality. Anthropic's daily OpenRouter usage surged from around 0.6 trillion tokens per day through mid-September to above 2.2 trillion by October 24, after rolling out Opus 5.5, Sonnet 5.5, and Haiku 5.5.
An analysis by Max Woolf, published October 26, finds AI model API prices declining rapidly and links the drop to growing competition from open-weight models.
Vercel's AI Gateway model leaderboard shows 'Jev' leading enterprise model preference at 16.4% and reach at 23.2% among gateway teams, with Anthropic leading text generation and OpenAI producing more than half of all images. Rohan Paul published results for an RL post-trained Qwen3-27B showing hallucination rate dropping from 30.4% to 15.1%, with a compressed variant fitting on a single 24 GB GPU.
NYU professor Aswath Damodaran argued the $10-15 trillion AI market is only achievable if AI replaces workers, which he calls economically and socially destructive.
bunq reports Finn on open-weight models via Nebius costs 60% of closed models per answer with matching quality
Bank of America found AI capex running 50% above forecasts while AI revenue runs only 20% above; projects capex reaching $1.5T by 2028.
Jefferies strategist Chris Wood published base case for 'massive capital destruction in the US' from AI buildout
NVIDIA announced at GTC Berlin expansion of open-model portfolio for European sovereign AI
Bain report projects AI must generate $6 trillion annually by 2031, with a $4.2 trillion gap versus identifiable revenue sources
Vercel's AI Gateway data through August 22, 2026, puts open-weight models at 62% of enterprise token volume but under 9% of spend, and clarifies that an earlier 29% figure covered only Chinese-built models, a subset of the open-weight total.
In July, Anthropic collected 65.1% of all gateway spending on 29.8% of volume, at 4.4 times the average price, while closed models held two-thirds of dollars. Z.ai's GLM 5.2 grew 27x in daily tokens and 80x in customer count in its first week, the fastest adoption Vercel tracked in 2026.
Vercel AI Gateway data shows open-weight models at 29% of tokens in June, up from 11% in April
BackgroundFrontier models, open models, and how they are compared
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.