Korva is a scroll of the internet’s best ideas.
Microsoft was the first yes
The conviction
The conviction came with GPT-4. The model was smart enough that OpenAI could see an approach that would work for reasoning, and reasoning working would bring about what is now called agents: systems that go do hugely valuable pieces of economic work.
The skeptics
Almost everyone called the plan totally crazy and impossible. No industry has ever moved like this, there are booms and busts, demand won't go up in a straight line, and the whole thing was reckless. It felt a lot like fundraising for an early-stage startup, where most people say no and one or two yeses are all it takes.
The first yeses
Microsoft was the first yes. Oracle then became a very big yes on the cloud side, and Nvidia has been a tremendous partner.
Why does reasoning need so much more compute?
Reasoning models spend compute while answering, generating long chains of intermediate steps before they commit to a final response. Each generated token requires a full pass through the model, so longer thinking costs proportionally more. Millions of users each running such chains multiply the demand further. OpenAI's o1 models, announced in September 2024, reported accuracy that rose as more compute went into thinking time.
Why do chip and data center projects take years?
A leading-edge chip factory, called a fab, costs billions of dollars and typically takes several years to build and bring online. Large data center sites also need new power connections, which can take years to arrange. The most advanced chips depend on specialized suppliers. ASML of the Netherlands is the only company that makes the extreme ultraviolet lithography machines used for them.
Why did earlier infrastructure booms overbuild?
In the late 1990s, telecom companies laid vast amounts of fiber-optic cable in expectation of explosive internet traffic. When the dot-com crash arrived in 2000 and 2001, much of that capacity sat unused and became known as dark fiber. Several large carriers filed for bankruptcy. The surplus later made bandwidth cheap for the companies that came next.
More from Invest Like the Best with Patrick O'Shaughnessy
