What the AI Midgame Looks Like

In chess, there’s a concept called the early game, the midgame (or middle game), and the endgame.

We just crossed the boundary into the middle game with the creation and subsequent restriction of Anthropic’s Mythos/Fable class models, and OpenAI’s GPT 5.6 Sol, Terra, and Luna models.

Let’s unpack what this does and does not mean.

First, we need to allay any fears that the US government requesting (or forcing) these frontier labs into pulling or holding back their latest models means that we are heading for a dystopian outcome.

Many people fear that this is a prelude to banning open source AI or dramatically restricting frontier AI, creating a bifurcated landscape.

Official US policy, legal analysis, and technical experts all say no, we are not heading towards a high-control regime.

Instead, we are now moving into a phase transition where “OpenAI and Anthropic cannot just drop a frontier model on the market sight unseen on any given Thursday.”

The new paradigm is “Defenders First” and the nuance should dramatically alter your interpretation of the government’s crackdown on Mythos and GPT 5.6.

Defenders First means, exactly as you’d expect, frontier AI labs are expected (or required) to give privileged access to key strategic partners first. That means certain government departments, technology partners, and cyber security firms so that they know what’s coming. This is just the new normal.

It does not mean that the government is on the verge of cracking down on open source AI models or going to pull up the ladder permanently.

The United States explicitly wants an accelerationist policy, which means they want the market to get their hands on frontier AI. The only caveat is that national security trumps market expediency, and even then, only temporarily.

There is one huge caveat to this whole thing, though, which is “what about the rest of the world?”

The US is pursuing an explicitly hegemonic strategy with what is being called Pax Silica. The idea is that a US-led coalition of allies should work to secure the full AI stack, lock stock and barrel, and that US allies across Europe, Asia, and elsewhere, should all default to American AI models.

That is one polarity.

The other polarity is what the UN is calling “Sovereign AI” whereby nations should create indigenous capacity for the full AI stack (chips, data, talent, data centers, models). The US is adamantly opposed to this for obvious strategic reasons. If you control the stack, you have unilateral discretion, which translates to leverage and coercive power.

Neither side is getting their wish.

Pax Silica is impossible because the US does not own or control all the bottlenecks, namely, EUV lithography. It is true that the US has the lion’s share of hyperscalers, which counts for a lot. But another dimension that proves that the US cannot gain unilateral control over AI is the fact that China is successfully competing with the US, and it’s almost entirely on indigenous capacity. They have not yet reached parity on technologies like EUV, but they have already crossed the threshold of “good enough to scale” while they catch up on the fundamentals.

Conversely, Sovereign AI is impossible because most nations simply cannot afford the chip, compute, and energy infrastructure required to host fully indigenous capacity. Even though political willpower around the globe is rising, since Europe and Asia do not want to be arbitrarily cut off from critical AI infrastructure at the whims of the US government, political desire alone is insufficient.

Nations such as Japan and India are actively developing onshore AI capabilities, but even so, they will be dependent upon Western nations or companies at some point in the supply chain. Germany and other European nations see US-dependence on AI as a threat to national security.

The inevitable outcome, given where things stand today, is a fragmented scheme of control and geopolitical chess where AI is concerned. China is going to continue to release open source AI to drive adoption of Chinese technology and to undercut Silicon Valley. The US is going to continue to try and monopolize the supply chain. But the cat is out of the bag on AI.

Regulations and export controls are migrating “from chip to cloud.”

The US government has realized that chip export controls aren’t working, and don’t really help. So they are pivoting to controlling access of frontier AI models at the API level. It remains to be seen if this will work as intended, or if it will simply aggravate global allies, such as Japan, Germany, and Australia more.

But there’s another trend on the rise, which I personally think most people have not priced in, and that is the dominance of local AI.

Several whitepapers and business trends have all converged on one very important observation: local AI is very nearly good enough to do most tasks. Right now, “cloud is the most defensible moat” is treated as an axiomatic, unassailable belief.

Sure, you need a hyperscaler to train a model. But more and more models are able to run locally, and do real meaningful work with consumer graphics cards. That trend will only continue. We’re already passing the intelligence optimum for many tasks, even coding, as small, local models begin to saturate coding benchmarks.

It’s true that a data center with a million GPUs will be able to run more AIs in parallel, and larger models, but we’re fast approaching a new post-scaling regime where instead of “bigger and smarter” we all want “faster and lighter.”

The market is already selecting for cheaper, lighter, faster, more efficient models because the level of intelligence is “good enough.”

The simplest way to understand this is through several first-principles lenses.

The mantra that I developed here is “mass, distance, time, and energy.”

All economic activities suffer from regress to one (or more) of these bottlenecks. You could have an AI model with an IQ of 3000 that costs three cents per billion tokens. But that amount of intelligence won’t immediately make concrete dry faster, shipping lanes go faster, or shrink the distances that goods have to travel. Physics sets the cadence of most activity, and eventually, you’re going to have a “country of geniuses” sitting around bored with nothing to do.

There is, however, a major exception. What I just outlined is the limitations of excessive intelligence at the operational edge. However, excess intelligence pays dividends strategically, over the long run. Put it this way, America got the bomb because we had Einstein and Oppenheimer on the job. Discontinuous leaps do come from excess intelligence.

This means that yes, at the long-term, strategic level “more intelligence is always better” can serve as a good heuristic. But at the short term, microeconomic level, “more intelligence” does not always translate to better outcomes. What you really want is cheap, fast, “good enough” intelligence that is fit for purpose. Cost-optimized intelligence is the next big target, and local AI is already starting to satisfy that target.

Even today, frontier models are generally overkill for all but the most challenging tasks.

And this changes the Pax Silica vs Sovereign AI picture.

The reason is simple: within 5 years, everyone with a decent laptop, let alone a decent data center, will be able to run more than enough AI to saturate their economic needs. Yes, there will always be uses for excess intelligence at the margin, but most cases will not be rate-limited by intelligence.

And you won’t need gigantic data centers to do that, either.

We can call the current technological, political, and economic phase the messy middle.

It’s messy because the age of “just scale” is over, as is the age of “release week” going off like a shockwave across the news cycles. Furthermore, no one team, group, nation, or company is going to get all their wishes met.

Data centers will continue to be built out, even as some states resist them. No one has an unambiguous victory.

China will continue to catch up even as the US works to secure hegemonic power. Neither will enjoy a clean victory over the other.