And Cloudflare wants to tax the robots! #
- Welcome to *. Cautious Optimism, a newsletter on tech, business, and power. Modestly upbeat
Friday. The United States’ economy stumbled in July, shedding 23,000 jobs. The market had expected jobs to improve by 83,000.
But what are stocks doing in response? Rising! Investors reckon a weak labor market will prevent the American central bank from raising interest rates to combat persistently above-target inflation. Small gifts, indeed.
What’s worse: May and June payroll growth was revised down by 103,000 jobs, dramatically lowering employment growth in the months directly preceding July’s net decline. Not good.
The stock market is not the economy. Tech stocks inform the stock market’s movement to a large extent today, so, by transitive property, tech companies are not the economy.
Now that we’ve established that, let’s discuss the end of Chinese AI labs’ low-cost regime; Cloudflare’s ambitions; and Google becoming a neocloud. Afterward, we’ll get into both what’s Trending Up and Trending Down. To work! — Alex
Chinese AI gets expensive #
The launch of Moonshot’s Kimi K3 made the world sit up and take notice. Here was a *very strong *Chinese AI model, and its maker intended to share its weights publicly. Were American AI labs about to get mogged by a far smaller company?
No, for two reasons. First, Kimi K3 is not that cheap, with headline token prices of $3 and $14 per million input and output tokens, respectively, which places it between what GPT-5.6 Terra (on a per-task basis) and GPT-5.6 Sol cost — makes sense, in a way, as they bracket K3’s capabilities.
Secondly, when K3’s weights were released, Moonshot added a clause demanding a commercial relationship to serve its inference — neoclouds have to pay a reported 30% cut to Moonshot.
In other words, Kimi K3 rips, but it tastes more like an American AI model than a classic model from the East: Not (entirely) open and not shockingly cheap.
Indeed, when we look at Artificial Analysis’ cost per task index, we can see two current-generation Chinese AI models (K3 and Qwen3.8 Max) near the top of the expense chart:
Look at the left of the chart, however, and you’ll find GPT-5.6 Luna, which remains an outlier for its cost/performance balance. You also see DeepSeek leading that end of the chart.
Will Chinese AI models hold on to their cost advantage? Perhaps not. DeepSeek customers ran into the following note in their API consoles this week:
That’s bad news, unless you are an OpenAI shareholder. If DeepSeek’s price hike proves material, and other Chinese AI labs like Moonshot are forced to charge rates closer to American prices to deliver near-frontier performance, the expected cost advantage of using China-built AI models will decline. Perhaps sharply.
I doubt I’m overstating the case. Reuters reported today that Alibaba intends to charge third-party inference providers to serve its Qwen3.8-Max model. How long until other Chinese AI labs stop leaving money on the table?
- A few weeks ago, OpenAI and Anthropic were washed. How quickly things change.
- As Chinese AI labs raise more money ( DeepSeek is raising billions at present), pressure to generate revenue at attractive gross margins will increase. And prices will rise.
Cloudflare wants to be the next Google
Cloudflare is betting big on agents. Its “Workers” platform was a key driver of growth in its second quarter, and paired its new Cloudflare.pay service (wallets and identity for agents) the company thinks it could shake up the Internet with simple agent-led micropayments.
Here’s CEO Matthew Prince discussing the economic challenges (and solutions) that agents offer the modern web:
Somebody has to pay for the bandwidth, somebody has to pay for the server, somebody has to pay for the people doing the work to create the content. I think that the business model of the Internet for the last 27 years has been largely defined by advertising and really defined by Google. I think the business model of the next 27 years of the Internet is going to be very different, and there’s no company in a better position to define what it looks like than Cloudflare.
*If *agentic traffic continues to scale as it has — Cloudflare said the majority of traffic “flowing across [its] network was not human” earlier this year — and if those agents need to pay for access to content, then a company that offers both a strong agentic platform and a simple way to allow those agents to buy access *could *become a primary broker of future Internet traffic.
Just like Google is today.
Google, the world’s largest neocloud
Subscribe to Cautious Optimism to unlock the rest.
Become a paying subscriber of Cautious Optimism to get access to this post and other subscriber-only content.