The Frontier Comes to You

TL;DR

• Kimi K3 shipped last week at 30% of frontier prices. A near-frontier model from a Chinese lab, with open weights promised within days. The model isn’t the real story. The rate of change is.
• AI efficiency doubles every eight months. That’s measured, not marketing. Chips take two years to double. The Moore’s Law of AI lives in the software.
• You never catch the frontier. It comes to you. A single consumer GPU today runs models that match, at least on benchmarks, the frontier of 6 to 12 months ago.
• The full experience is about three leaps behind that. Matching a benchmark is not the same as running well. Today’s best model quality, at real speed on a few thousand dollars of hardware, is roughly two to four years out.
• The weights are everything. An API can be repriced, restricted, or retired. A weights file on your disk cannot. Washington is already moving on Chinese models. Archive the open models while they’re freely available.

---

Another Tooth on the Gear

Last Wednesday, Moonshot AI released Kimi K3. It’s a 2.8 trillion parameter model that scores within a few points of Claude Fable 5 on independent benchmarks, and it costs $3 per million input tokens and $15 per million output. Fable costs $10 and $50. So the model sitting just behind the frontier now rents for less than a third of the frontier’s price.

The weights aren’t public yet. Moonshot promised them by July 27, license terms to be announced. And I should be honest about the rest of it too. K3 is near the frontier, not at it. Its “2.5x scaling efficiency” claim is the company’s own number, still waiting on a technical report. And in a small irony, K3 is actually a price increase. Moonshot’s previous model cost about a quarter as much. The era of rock-bottom Chinese pricing may be ending even as the models get good.

None of that dims the story for me, because K3 was never the story. It’s one tooth on a gear that has been turning at a steady, documented, frankly absurd rate for years. This post is about the gear.

The Question I Left Open

Last month I wrote Renting the Frontier, about what it would cost to run my own coding model and why the real issue was control, not money. I ended it with a loose thread. The thing that was clearly uneconomical last year is merely expensive this year, and next year it’s something else.

That sentence was a feeling. I wanted the actual number. How fast does “uneconomical” become “expensive” become “why not”? So I went and dug through the research, most of it from Epoch AI and a team at MIT, and the answer is more concrete than I expected.

The compute needed to reach a fixed level of AI capability halves roughly every eight months. That’s from an analysis of 231 language models across a decade. For comparison, Moore’s Law doubled chip density about every two years, and it reshaped the entire world economy. The software side of AI is moving three times faster than the most famous exponential in the history of technology.

Prices follow. The cost to buy a fixed level of capability through an API has been falling between 9x and 900x per year depending on the task. GPT-4 level performance on PhD-science questions got 40 times cheaper per year. And the pace is accelerating. Restricted to data after early 2024, the median decline jumps from around 50x to around 200x per year.

Hardware is the slow lane, and that surprised me. GPU performance per dollar improves about 37% a year. Memory bandwidth, the thing that actually decides how fast a local model talks, grows about 28% a year. Almost none of this revolution is coming from the chips. The gains live in algorithms, quantization, sparsity, and better serving software. Which means they arrive on hardware you already own. NVIDIA made that point for me with their little DGX Spark desktop box, reporting inference speedups of around 2.5x within months of launch without touching the hardware. One leap, delivered as a software update.

The Conveyor Belt

Here’s the part that changed how I think about all of this.

You will never self-host the frontier. I want to say that plainly because I spent a while resisting it. The frontier model of the moment always needs a datacenter. K3, even after its weights open up, wants 64 accelerators to serve properly. Nobody is running that in a home office, and by the time you could, there will be a new frontier that needs even more. The researchers tracking this found that while fixed capability gets cheap fast, the cost of running the newest frontier model actually rises every year. The target moves.

But the frontier of a year ago? That comes to you. Epoch measured it directly. The best open models trail the closed frontier by about four months. Models you can run on a single consumer GPU trail it by 6 to 12 months, at least on benchmarks.

I want to be careful with that claim, because last month I put the practical ceiling for home hardware around Sonnet-class and said benchmarks undercount the slow grind of real agentic work. I still believe that. Both things are true. The paper gap closes first, and the felt gap follows it down with a delay of its own.

So the right mental model isn’t a race you’re losing. It’s a conveyor belt. The labs push the frontier forward, and six to twelve months later, that same capability rolls off the belt and onto your desk. It has done this every year for years. You don’t have to chase it. You have to be standing there when it arrives.

What does the belt deliver in practice? Fable-quality on a $2,000 to $5,000 box needs something like three more big efficiency leaps, the kind Moonshot is claiming with K3, the kind NVIDIA shipped to the Spark. The measured rate says the field produces about one of those a year. Call it two to four years, with the hedge that the last stretch (real speed, honest quality, not just a benchmark number) always takes longer than the napkin math says. No moonshot required, just the belt doing what it has always done.

And I’d argue you don’t even need to wait for Fable-quality. The models rolling off the belt right now, the ones a Mac Studio or a used server can hold today, are already past the frontier of early 2025. Ask yourself honestly what your work actually required in early 2025. For most of what a small shop does day to day, the answer is already on the belt.

The Weights Are Everything

Now the part I actually care about, and the reason the July 27 date matters more than the benchmark scores.

An API is a relationship. A weights file is property. Everything I use through an API can be repriced, rate-limited, deprecated, or quietly changed under me, and over the past two years all of those things have happened. I’ve written before about models changing under my feet mid-project. These companies have not proven to be the kind of consistent partners you build a decades-long business on top of. I don’t say that with anger. Their incentives point at the frontier, not at my stability. But I have to plan around it.

Weights are the one artifact in this entire industry that you can truly own. A file on your disk cannot be revoked, cannot be migrated forward on someone else’s schedule, and cannot change its behavior overnight. If every AI lab on earth turned their APIs off tomorrow, a downloaded DeepSeek or Qwen or Kimi model would keep working, at exactly the quality it had, forever.

And the window matters, because access is getting political from both directions at once. Washington has spent the spring warning about Chinese models, and it’s now moving. Congressional committees are probing American companies that use models like Kimi. Federal procurement bans are on the table. Meanwhile the Chinese labs are raising prices and there are reports Beijing may restrict what its labs release overseas. I have no idea where any of that lands. What I do know is that an API can be turned off in an afternoon, and a file on a disk you own cannot. That difference is the entire point.

So here’s my plain recommendation. If you have the disk space, and following whatever rules apply to you, download the significant open models as they’re released and keep them. The big open models run from a few hundred gigabytes quantized to over a terabyte at full size. A 4TB drive costs less than a nice dinner out and holds a meaningful archive. That drive is the cheapest insurance policy in software right now. If K3’s weights land as promised on the 27th, I’ll be archiving them, and I’d suggest you do the same.

Standing at the End of the Belt

Where does this leave a small shop like mine?

Not off the frontier APIs, at least not yet. I said in the last post that I hadn’t moved my day-to-day off the frontier, and that’s still true. The current frontier is genuinely better at the hardest work, and I’d be lying if I said otherwise.

But the long game has come into focus for me. The endgame is autonomy. A business that plans in years can’t have its core capability depend on a vendor relationship that changes monthly. And for the first time in this industry’s short life, the trend lines say autonomy is actually coming. The efficiency curve is faster than Moore’s Law. The belt reliably delivers last year’s frontier to hardware normal people can buy. The open models sit only about four months back (a gap that has drifted wider by a month or so over three years, so I’m watching it, not assuming it). Every year, the price of a fixed level of capability collapses, and the part of my workflow I could own instead of rent gets bigger.

So my strategy is simple, and it costs almost nothing. Keep renting the frontier while it’s clearly worth it. Archive every meaningful set of weights as they ship. Test the local option a couple times a year so the exit is real and not theoretical. And let the conveyor belt do the heavy lifting, because it’s powered by the entire field’s research and it doesn’t care which lab is winning.

Eighteen months ago I would have told you self-hosting serious AI was a hobby for people with server racks. Now I think the more accurate framing is that everyone self-hosts eventually, just on a delay. The frontier comes to you. The only question is whether you’ll be ready to receive it.

I plan to be standing at the end of the belt.

---

David Kerr is the founder of Kerrberry Systems. He builds custom software for businesses that want to own their systems, not rent someone else’s. Find him on LinkedIn or GitHub.