Compute is centralizing. Your personal AI never needed it.

TL;DR

Dylan Patel told Dwarkesh Patel that OpenAI and Anthropic are on track to control most of the world's usable compute by the end of 2028, and neither of them could name a force that stops it. The math is sound. The conclusion people are drawing from it, that a personal AI on hardware you own is finished, is not. Every force they name belongs to one race: who gets to train the next best model in the world. Your agent does not train anything. It runs, and by SemiAnalysis's own measurement the cost of running a given level of capability keeps falling.

The strongest public argument against personal AI this month was not written by a critic. It came from the person who tracks where every megawatt of AI compute in the world goes.

The case, at full strength

On the Dwarkesh Podcast, SemiAnalysis founder Dylan Patel laid out the ladder. About a third of the compute coming online this year serves OpenAI and Anthropic. Next year, on contracts already signed, they take 40 to 50 percent. By the end of 2027, half of all new compute in the world is theirs.

They can outbid everyone because they turn a megawatt into more money than anyone else can. Compute costs 10 to 15 million dollars per megawatt. Anthropic's revenue on one has reached 50 million. Behind that sits a little over a trillion dollars of AI capital spending this year and more than two trillion by 2028, a large share of it borrowed.

Patel's conclusion, in his own words: by late 2028, "you've got them just controlling most of the usable flops in the world on their own."

Then the episode does something unusual. It ends on the question it could not answer.

Dwarkesh named the same gap in his own write-up: whether anything can counter "the economies of scale in training, the scarcity of compute, and eventually continual learning and RSI." Two of the best-informed people on compute went looking for the counterforce and did not find one.

Read that list again

Economies of scale in training. Scarcity of compute. Continual learning. RSI.

Every item is about the same question: who gets to build the next best model in the world. That is a race to train.

Your agent does not train anything.

Its job is to run. Hold what you told it. Remember your sizes, your budget, the date the thing has to arrive. Keep your credentials. Call the tools. Do the errand while you are asleep. That work needs a model that is good enough at following instructions and using tools. It has never needed the best model in the world, and it never will. Nobody needs a frontier model to remember their shoe size.

So the forecast and the fear are about two different things. The forecast is about training. The fear is about running. They got attached to each other and nobody separated them.

Their own numbers say good enough keeps getting cheaper

The evidence for that is not ours. It is theirs.

Four days before the episode, Patel's own publication measured how fast open models catch up. Their finding: with each generation, "open-source models take half as long to catch up to the first closed-source model of the era." The reasoning era gap closed in 8.5 months. In the agentic era, Kimi K2.6 passed Opus 4.5 in 4.8 months and GLM-5.2 cleared GPT-5.2 in 6.

Running those models is already cheap enough to be a business. Patel again, describing the commodity market plainly: "anyone can make money off of $10-15 million per megawatt compute today. I kid you not, it's not that hard. Go get a GB300 rack, go download the Kimi weights, go download vLLM or SGLang, set it up." Put it on OpenRouter, he says, and "you'll start generating more revenue than you're paying for the compute."

And the labs are walking away from that market on purpose. Patel calls it his non-consensus view: "The labs are going to allocate less and less compute to inference over time." They want the compute for training. The thing your agent needs is the thing they are giving up.

Put the three together. The frontier gets more concentrated and more expensive. The level your agent actually needs gets cheaper every year. Both are true at once, and the people making the centralization case are the ones who measured the second half.

The 4GPTs read: "compute is centralizing" is a fact about the race to train, not a fact about every use of a computer. Personal AI sits on the other side of that line, which is why the forces barrelling toward the center never turn to face it.

What would prove this wrong

Two things, and they deserve to be said plainly.

The frontier stays closed and ahead. The halving clock measures how long open models take to reach a level the leaders have already left. SemiAnalysis say it about their own habits: "we still prefer using Fable at SemiAnalysis for our day to day work." The claim here is that your agent's job does not require the best model in the world, not that the gap disappears.

Continual learning could join the two markets back together. If models improve mainly by learning from millions of deployments, then a personal agent's capability starts depending on whoever owns the biggest deployment, and the argument above weakens. Patel names this force directly. It has not arrived yet. If it does, this piece is wrong, and it will say so here.

Where to start

Nothing in that episode is a reason to stop owning your stack. The race it describes is real, and you were never running in it.

If you have not built the stack yet, your personal AI stack, from zero walks the install order. The shopping capability is SIL: built on UCP, runs on OpenClaw, open source under Apache-2.0, and it holds your requirement on your side of the table.

FAQ