In the days since releasing my last article, I’ve heard from a lot of people. There’s so much confusion around AI.
A large amount of the disconnect, I think, comes from how people use AI in their lives right now. Plenty of people don’t use it that much. Those who do often aren’t exposed to the frontier of publicly available AI, which is still very expensive—especially inside enterprise contracts—and delivered through systems that are still quite crude. So it’s understandably hard to see the harm that industry insiders claim is coming. And it’s reasonable to question whether it’s a marketing stunt. Such is life in a capitalist society.
But perspective is worth considering.
Even further from where literally 99% of people use and interact with AI is the frontier of the technology, which is being developed inside model labs across the world, often under strict confidentiality. These people have access to models most enterprises and individuals can’t afford: models the public will call “frontier” for six months after release, themselves often six months after initial training and lab access.
These people are starting to speak up. They are saying that we cannot afford to keep racing.
I would argue that the next-closest people to where the frontier of AI resides are those building extensively with AI on the subsidized $200 OpenAI and Anthropic plans—probably another 1–2% of people. This subsidized access enables prosumers to build quite effectively with expensive models long before that level of artificial intelligence is available to those paying little or nothing for AI, because the subsidy affords them what would cost thousands of dollars through enterprise plans.
In the 30 days ending September 14, my usage through one $200 Anthropic plan and one $200 Codex plan—$400 a month combined—amounted to an estimated $20,173 at API list prices, more than 50 times what the plans cost.
Usage breakdown: $20,173 at API prices
Activity Daily token traffic
30d
Economics Estimated cost by model
API list prices
- GPT-6 Astra $9,313 46.2%
- GPT-5.6 Sol $6,898 34.2%
- Claude 5.1 Fable $2,900 14.4%
- Claude 5 Fable $874 4.3%
- GPT-5.6 Terra $174 0.9%
- Other $14 0.1%
And the harness matters—or, to put it in more layman’s terms, how you access these models matters. Good applications and harnesses around a model can reduce the cost per task dramatically when developing at the forefront of AI, where the most expensive models are involved. My own AI usage is through applications I’ve tailored extensively for this purpose. It’s part of the equation behind Cursor’s roughly $60 billion sale to SpaceX, but for the purposes of relevance to this article, I’ll leave it at this: a better harness gives you more exposure to what’s going to be possible and affordable in roughly six months’ time.
These people are starting to speak up. I’m starting to speak up. I don’t like the look of this race.
Despite the recent developments in the news, today, across the globe, labs are still racing. Incentives are not aligned toward slowing down the training of AI models. And nobody is even considering how one might slow down advancements in harness or hardware engineering that accelerate AI models. Not in the slightest. Not even a little bit.
There have been a lot of words said, but actionability remains zero. There might be more safety monitoring coming… maybe.
Here’s the current state of play as I see it. The United States and China are the two nation-states contending in this race. Europe is… not really involved.
Right now, the US is “ahead” and China is closing “fast”. That’s at least probably the most commonly accepted public narrative.
That’s because the US has maintained control of the hardware needed to train AI, with a few asterisks. Right now, the US is maybe 2–4 months ahead in raw frontier-model intelligence.
BUT AI being developed in China is way more affordable. So much more affordable that I cannot imagine it isn’t making US AI companies extremely uncomfortable. There’s really no economic reason to develop with US AI companies, I would argue, for most tasks. In that sense, China is very, very much ahead. DeepSeek’s published prices make the same point.
China is also arguably ahead on harness architecture, though it’s unclear what exists inside US frontier AI companies.
And so, the real lead the US has is in frontier-model intelligence. That’s the thing that leads to RSI, AGI, or ASI most directly, and what probably snuffs us if it takes off—potentially if it’s created in any capacity, almost certainly if it isn’t created extremely safely. But that’s the real lead the United States has.
So none of these labs can actually afford to slow down in the competitive landscapes they are operating in. They simply can’t. LEGALLY, it isn’t even clear that they can coordinate to do so without government action: an agreement among competitors to restrict output can itself create antitrust problems. Which takes me to what they all keep talking about.
“Policy”
It’s a polite word for an impolite time. Right now, AI labs are jumping over each other to ask for government regulation. Whether frontier AI labs genuinely want policy for safety—and not instead for regulatory capture and evasion of commitments—remains unclear. I think healthy skepticism is warranted, but it should not detract from a clear-eyed assessment of the risks.
Regardless, the usual state of policy, even at the best of times, is slow. You know the government that exists in the United States right now. No matter what side of the aisle you reside on, I don’t imagine you think the US does anything critically important with urgency.
We haven’t seen a senator for three months. Midterm jockeying is fully upon us. The president is signaling race (signaling hard). It’s probably the worst time in history to have to make a quick, long-term strategic decision, and it’s exacerbated by the fact that a huge amount of US economic growth—measured quarterly—is riding on AI.
Yep, the US economy is totally strapped in. And you know what our president and any political group knows is a death knell? Stock market collapse. Slow down and lose the only remaining lead you have? I think not.
For these reasons, I’d be surprised if we saw US-led policy regulating or slowing down AI development.
Hence, we are still racing.