“I am paying for tokens that create no value.”
That was Palantir CEO Alex Karp on CNBC, describing what he says he hears from frustrated business leaders. If you’ve used tools like Claude code or OpenAI’s Codex, you may have experienced it yourself.
Karp is talking his book. Palantir sells the software layer that helps companies turn foundation models into useful enterprise applications. But his criticism is still valid.
For the past two years, Silicon Valley has treated token consumption as a sign of progress. Companies encouraged employees by giving them all subscriptions to tools that were meant to increase productivity, and encouraged them to use them heavily. In Shopify for example, they ranked employees according to their use of AI for their role in their performance review. More prompts, more agents and more tokens per employee supposedly meant a company was becoming AI native. But tokens are not an output. They are an input cost.
A company can consume ten times more tokens without producing ten times more software, resolving ten times more customer requests or generating ten times more revenue. Sometimes the model is simply thinking longer, repeating work or using an expensive frontier model for a task that a cheaper model could handle.
That calculation is already changing. The Financial Times recently reported that companies including DoorDash, Airbnb and Siemens are using Chinese AI models, attracted by their lower prices, improving performance and open weights. In some cases, the models can be up to 60 times cheaper than their American competitors.
This would have sounded unlikely a year ago. US companies were expected to build on OpenAI, Anthropic or Google. Instead, Chinese models from Alibaba’s Qwen, Moonshot’s Kimi, DeepSeek and Z.ai have become serious options, particularly for coding and agentic tasks. I wrote about the trend in November last year.
Cursor offers a good example. Its Composer 2 coding model was built by taking Moonshot’s open Kimi K2.5 model and training it further using Cursor’s own coding data and reinforcement-learning environment. Cursor says the result is competitive with leading frontier models at a much lower inference cost.
is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5. (source)
This may be the more important shift. The best model for a startup is not necessarily the model at the top of a general benchmark. It may be a cheaper open model that the company can adapt around its own users, data and workflow.
Until this week, Chinese labs had a clear advantage in that part of the market. Their open models were often close enough to the frontier, far cheaper and easier to customise.
This week (July 15), Mira Murati’s Thinking Machines Lab launched Inkling, an American open-weight model trained from scratch. The company is unusually direct about its limitations:
“Inkling is not the strongest overall model available today.”
It trails the best closed models and several Chinese models on important benchmarks. That might make the launch look underwhelming. I think it may be a smart move.
Thinking Machines is not trying to win only by producing the highest benchmark score. Inkling is designed as a broad base model that companies can customise. It handles text, images and audio, is available for fine-tuning and can vary how much effort it spends on a task. Simple work can use fewer tokens; difficult work can be given more reasoning time.
That is closer to what the market needs. Most businesses do not require the world’s smartest model for every step of every workflow. They need a model that is reliable enough, fast enough and cheap enough to use at scale. They may also want open weights, control over deployment and an American alternative to Qwen or Kimi.
The next wave of AI companies will not be built by tokenmaxxing
They will use frontier models where maximum intelligence matters, cheaper models for routine work and specialised models for specific workflows. They will route tasks between models and measure the cost of reaching a successful outcome, not the number of tokens generated along the way.
For founders and investors, the useful questions are becoming simpler:
Did the task get completed? Was the result correct? What did it cost? Does the product become more or less profitable as usage grows?
Tokens are the fuel. They should never have been the destination.
He is a former General Partner at Google Ventures (GV) in Europe, former head of Google for Entrepreneurs in Europe, and founding head of Campus London, Google's first startup hub. Eze writes on Israeli tech, venture capital, artificial intelligence, and founder strategy.
He is also the founder of Techbikers, a nonprofit that brings together the startup ecosystem on cycling challenges in support of Room to Read.
- Tokenmaxxing Was the Wrong Metric - July 16, 2026
- The AI Shovel Paradox: Why Israel’s AI Future Lies in Software and Silicon, Not Server Farms - July 15, 2026
- Weekly Firgun Newsletter – July 10 2026 - July 10, 2026
