AI's Distribution Race Is Failing on Both Ends

This week, OpenAI removed text chat rate limits for its free and Go tiers and upgraded the default model to GPT-5.6 Luna — a move to let anyone in, no friction, no gate ^[1]. Four months ago, Anthropic blocked third-party tools like OpenClaw from accessing its Claude subscriptions and began throttling session limits during peak hours — a move to keep people out, or at least to make them pay for the privilege ^[2][3]. Two of the most-funded AI companies on earth, running opposite distribution strategies, and neither strategy is winning the distribution race. OpenAI's open-door bet is demonstrably failing: the company missed both its internal revenue targets and its goal of one billion weekly active users, and its CFO warned it may struggle to fund $600 billion in compute contracts without accelerated growth ^[4]. Anthropic's walled-garden approach caps its reach by design — blocking the very power users who would drive adoption, pushing them toward competitors or toward the paid API tier that most enterprises are now scrutinizing ^[2][3]. The contradiction is not a quirk of two companies making different bets. It is the structure of a race that cannot be won from either direction. Then the contest shifted to distribution. The logic was straightforward: models were becoming commodities, so the winner would be whoever put them in front of the most users. What followed was a race to give AI away. DeepSeek permanently cut V4-Pro prices by 75% in May, making it 12 to 19 times cheaper than equivalent OpenAI and Anthropic models ^[5]. In late July, OpenAI slashed GPT-5.6 Luna prices by 80% and Terra by 20%, a move widely read as a response to Chinese rivals. The company offered its own rationale.

I think we’ll have a lot of ways we can help people get more value for less spend. — Sam Altman

A Replit executive described the result.

GPT-5.6 Luna is the closest we've come to intelligence too cheap to meter. — Michele Catasta

^[6] Adobe went freemium, deferring price optimization to double its user base, with AI-first annual recurring revenue tripling to $500 million as the payoff metric ^[7]. Chinese open-weight models surpassed American ones in Hugging Face downloads, winning the distribution race through free model releases ^[8][9]. The problem is that giving AI away does not pay for itself. OpenAI's missed targets are the clearest signal: even the market leader cannot convert distribution reach into revenue at the scale required ^[4]. Chinese open-weight models have no profit path — Bloomberg warns the price war may prevent sector profitability for three years — yet firms continue cutting prices to capture market share ^[9]. Adobe is deferring monetization, betting that users acquired now will pay later, a wager that depends on a demand side that is already crumbling ^[7]. The demand side is the second failure, and it makes both distribution strategies moot. Corporate America's "tokenmaxxing" fad — employees burning tokens to signal productivity — has collapsed. Companies found token expenses doubling every two months without proportional value. Palantir's CEO was blunter.

As bills started to pile in, people realized that those new tools are quite expensive and you need to use them wisely. — Vincent Gusdorf

Enterprises are shifting to model routing, sending simple tasks to cheap models, and pivoting to Chinese open-source alternatives. Satya Nadella put the structural problem plainly.

In hindsight, that caution looks less like discipline and more like underestimating how fast demand would arrive. — OpenAI

The distribution victory is hollow if the product being distributed does not earn its keep ^[10]. One company is winning. Alphabet's Google Cloud revenue is up 63% to $20 billion, with a $460 billion backlog and $73 billion in free cash flow ^[11]. The reason is not that Alphabet built a better model than OpenAI or Anthropic. It is that Alphabet owned the distribution channels — Chrome, the world's most-used browser; Android, the world's most-used smartphone operating system; Google Cloud, already embedded in enterprise IT stacks — before the AI race began ^[12]. Its AI reaches users by default, not by buying its way in. Pichai calls it the "full stack approach," but the stack was already there. The models just had to be good enough. The capability race has not stopped. ByteDance is training a 10-trillion-parameter model to surpass Anthropic's estimated 8-trillion-parameter Mythos, with founder Zhang Yiming instructing employees to avoid distillation shortcuts ^[13]. OpenAI may need $600 billion in compute contracts to stay competitive ^[4]. Firmus Technologies just raised $2 billion to build AI infrastructure in Australia and Asia-Pacific, backed by Nvidia ^[14]. Capability is still necessary. It is just no longer sufficient. It is the entry fee, not the prize. The race the industry pivoted to is the one only a company that already owned the channels can finish.