Korva is a scroll of the internet’s best ideas.
Fall off the frontier, lose everything
Paying for the best
The two companies are on a hamster wheel. Most tokens will go to open or commodity intelligence, but a meaningful share of the market, perhaps 10, 20, or 30 percent, pays a huge premium for true frontier intelligence. Some customers need the best, and others want to know they have it, like a hedge fund that cannot risk a competitor having a better model.
That premium is why revenue keeps growing while token market share falls. Scale helps as well: something like 60 percent of the worldwide compute being added over the next year is going to these two companies, in the middle of a compute crunch, and they have been bringing their own prices down quickly.
A narrow lead
The frontier is maybe only six to 12 months ahead of the commodity portion of the market, depending on capability. A company that falls off the wheel for six months is in deep trouble, and that risk has to be priced in. Many of the open models are built by Chinese companies, who are not slowing down, so anything that slowed the two leaders would let the rest catch up.
Why pay a premium for the best model?
In competitive fields, a small edge can decide who wins, so buyers pay heavily for it. High-frequency trading offers a precedent: in 2010 a company called Spread Networks finished a straight fiber line between Chicago and New York at a cost reported at around 300 million dollars. The line shaved only a few milliseconds off the trip, and trading firms paid for access anyway.
How do open models catch up to leaders?
Most modern language models use the transformer design from Google researchers' 2017 paper Attention Is All You Need, so architecture ideas spread through published research. Researchers also move between labs, carrying techniques with them. Smaller models can be trained on text generated by larger ones, a method called distillation. Those routes shorten the gap, which is why the lead on a given capability tends to be measured in months.
Why does compute limit how fast AI grows?
Training and running large models depends on specialized accelerator chips, which are built with advanced packaging and high-bandwidth memory that few factories can produce. Data centers also need land, power connections, and cooling, and building one takes years rather than months. Orders for chips and power are placed long before the capacity arrives. That lag means the compute available next year was largely committed this year.
More from All-In with Chamath, Jason, Sacks & Friedberg
