The Weight of Empire: When Capital Outgrows Wisdom

I have inherited a few systems built by companies with more money than discipline, and the pattern is always the same. Nothing is optimized because nothing has to be. Every latency problem got solved by provisioning bigger instances, every architecture problem by hiring another team to work around the last one. The thing runs. It also costs several times what it should, and nobody noticed, because noticing was never anyone's job. Money is a design constraint like any other. Too much of it just points the wrong way.

I thought about that this morning, reading the funding headlines. OpenAI closed a $122 billion funding round. That number is correct. One hundred and twenty-two billion dollars, a single private raise larger than the GDP of about two-thirds of the countries on earth. The post-money valuation sits at $852 billion. An IPO is planned for the fourth quarter. Revenue is running at two billion per month, which sounds enormous until you start doing the arithmetic on what it costs to train and serve the models generating that revenue. The margins are a topic nobody at the celebratory dinner wants to bring up.

I do not think this is necessarily foolish. I have been wrong before about what markets will bear. But a raise is a statement about burn, and a raise this size is a very loud one. You do not take $122 billion because you found a cheap way to serve inference. You take it because your cost curve runs ahead of your revenue curve and you have decided to outrun the gap rather than close it. That is a legitimate strategy, and a bet that the gap is temporary. The question is not whether OpenAI can raise this money. They obviously can. The question is what the size of the raise says about the cost curve underneath, and whether that structure can hold the weight now being placed on top of it.

Morgan Stanley Sounds the Alarm Nobody Wants to Hear

While the champagne was being uncorked in San Francisco, Morgan Stanley published a report that the people uncorking it should probably read. Their analysts are telling institutional clients, plainly, that a massive leap in AI capability is imminent and that most organizations are not prepared for what it means.

The thesis is blunt. Scaling laws are holding. Throw ten times the compute at a large language model and you roughly double its effective intelligence. The April-to-June window of this year, which is to say right now, is when the effects become impossible to ignore. GPT-5.4 already scores at or above human-expert level on economically valuable tasks. The next generation of models will widen that gap further. Morgan Stanley's word for the consequence is deflationary, which is Wall Street's polite way of saying that a lot of people currently drawing salaries will need to find something else to do.

What strikes me is not the prediction itself. We have heard variations of this for two years. What strikes me is the source. Morgan Stanley does not traffic in blog-post futurism. They are telling wealth managers and pension funds to reposition. When that kind of institution starts using that kind of language, the abstraction phase is over. We are in the implementation phase, and the implementation phase is where things get genuinely difficult for people who were hoping the abstractions would stay abstract.

Thirty-Eight States Walk Into a Regulatory Bar

Here is the part that would be funny if I did not have to build against it.

The White House released a National Policy Framework for AI on March 20th. The framework recommends federal preemption of state AI laws. It wants a unified national approach. This is a reasonable thing to want in the same way that wanting everyone at a dinner party to agree on the music is reasonable. The problem is that thirty-eight states have already passed their own AI legislation, and several of those laws took effect on January 1st of this year. California alone has three major acts running. Texas has its own governance framework. The horse is not merely out of the barn. The horse has incorporated in Delaware and is filing amicus briefs.

The tension here is genuinely fascinating and I think it is going to define the next decade of technology policy in America. Washington wants to clear the runway for AI companies. The states want guardrails. Both impulses are defensible on their own terms. Together they produce a regulatory environment so fragmented that compliance becomes a competitive moat in itself, which is exactly how the largest companies prefer things, because they are the only ones who can afford the lawyers.

I build against this every week, so let me be concrete about the cost. You do not write one system and flip a per-state flag. You build for the strictest jurisdiction in your footprint and eat the overhead everywhere else, because maintaining divergent retention windows, disclosure text, and audit logging across a dozen regimes is how you get a compliance matrix nobody can reason about. So everyone quietly builds to California, and that lands hardest on teams with no legal department to tell them which of the thirty-eight even apply. Fractured rules do not produce lawlessness. They produce a system where the rules bind only those too small or too slow to choose where they stand.

The Model Race Enters a New Phase

Meanwhile, the technical work continues at a pace that makes the regulatory scramble look even more futile. This month brought Claude Mythos 5, ten trillion parameters aimed at cybersecurity and advanced coding. Google pushed out Gemini 3.1 with real-time voice and image analysis. But the announcement that caught my attention was quieter and, I suspect, more consequential: Google published a compression algorithm that cuts AI memory requirements by a factor of six.

That last item matters more than the headline models, and the reason is structural rather than spectacular. The bottleneck in AI deployment has been shifting for about a year now. It is no longer primarily a question of whether the models are good enough. They are good enough for a staggering range of tasks. The constraint is cost. Inference at scale is expensive. Serving millions of concurrent users on models with hundreds of billions of parameters requires hardware budgets that make the Pentagon blush. A six-fold reduction in memory overhead does not make a better model. It makes the existing models economically viable in contexts where they previously were not. That is a different kind of revolution, the boring kind, the kind that actually changes things.

The mechanism is the part that gets skipped, so here it is. Serving a transformer is memory-bound long before it is compute-bound. Weights and KV cache sit in VRAM, and the cache grows with every concurrent request and every token of context. Memory per request caps batch size, batch size sets tokens per second per GPU, and that number is the unit economics of an inference business. Six times less memory means more users on the same fleet and longer context at the same price. Nobody writes a headline about that. It is still the item on this month's list most likely to change what gets built.

Meanwhile, at the Moon

In news that has nothing and everything to do with AI, NASA is launching Artemis II today. Four astronauts. First crewed mission beyond low Earth orbit in over fifty years. Victor Glover becomes the first person of color, Christina Koch the first woman, and Jeremy Hansen the first non-American to travel beyond the neighborhood of our planet. The launch window opens this evening from Kennedy.

I mention it for the contrast in engineering culture, which I find more instructive than any valuation. Artemis II has a launch window, a crew, and no rollback. You cannot ship a rocket at eighty percent and patch it in the next sprint. Every subsystem carries margin somebody had to justify in a review, and the program is organized around a failure mode measured in lives rather than churn. Software does not work that way, and mostly it should not have to.

There is something clarifying about watching a rocket leave the atmosphere on the same morning you read about a company raising $122 billion to predict the next word in a sentence. Both are genuine achievements. But one of those programs has to prove it works before it runs, and the other finds out in production. We iterate because iteration is cheap, and when it stays cheap for long enough, you lose the ability to tell a system that works from a system that has merely not failed yet. That distinction is worth holding onto before we go back to arguing about valuations.

What Actually Holds

If there is a thread connecting this morning's headlines, it runs something like this: the infrastructure of artificial intelligence is consolidating at a speed that outpaces our capacity to govern it, fund it sustainably, or even fully understand what it is becoming. OpenAI's raise is a bet that scale will solve the problems that scale created. Morgan Stanley's warning is a bet that scale will break things faster than institutions can adapt. The states are legislating in real time. The models are compressing. The rockets are launching.

I have no forecast to offer here. I cannot tell you where the valuation lands, I cannot time the correction, and I have no idea which state passes what in June. Anyone selling you a number on those does not have one either. What I can control is the cost structure of what I build and the honesty of my own estimates. Build things whose costs do not require the capital markets to remain permanently irrational. Ship capabilities that function when the hype cycle turns. Design systems that survive contact with regulators who have not yet decided what the rules are.

The weight of this particular empire is considerable. The question, as always, is whether the foundation can carry it. I do not know the answer. Nobody does. But I know that the people asking the question are in a better position than the people who assume the answer is yes because the check cleared.

So here is what I actually do about it. Every agentic system I put into production carries one number: cost per completed task, measured rather than estimated. Then I ask what that number does if inference gets ten times more expensive, and if it gets ten times cheaper. Most of what I build survives the second case. A few do not survive the first, and I know which ones. That is not a prediction. It is knowing where my own foundation is thin, which is the only part of this I get a vote on.