Key Takeaways

  • OpenAI and Meta are buying every GPU they can find, according to Tae Kim, with executives at both companies saying they will take any chip available.
  • OpenAI's lead chip architect confirmed to Kim that the team now deploys autonomous AI agents to speed up chip design and silicon bring-up.
  • Tier-two neoclouds face a sharp capex squeeze over the next 12 months as Federal Reserve interest rate hikes and tighter credit dry up debt financing.
  • The push toward recursive self-improvement creates a compounding hardware loop: smarter models require exponentially more compute, forcing labs into custom silicon like Google TPUs and AWS Trainium.

The Neocloud Credit Crunch

Aggregate demand for compute looks bottomless, but the plumbing underneath it is cracking. Tae Kim, author of The NVIDIA Way, points out a growing split in the market. Frontier labs continue to hoard every accelerator they can secure, yet tier-two neoclouds that borrowed heavily to buy clusters face a harsh reality.

“On the margin I think the financing environment for neoclouds has become more difficult,” (play from 18:46) Kim explained. “And there's a lot of macro reasons for that with the Fed actually hiking rates and the financial conditions tightening.” (play from 18:51)

When interest rates hovered near zero, leasing GPUs on five-year payback periods with debt looked like free money. With higher rates and stricter lending standards, the math breaks. Frontier labs will keep funding their own clusters, but smaller cloud providers will struggle to finance their capex over the next year. If you depend on a discount neocloud for model training, their cost of capital is about to become your problem.

Agents Designing Their Own Silicon

The compute shortage has pushed companies beyond Nvidia order forms. Every major lab is building proprietary silicon. The surprise is how they are doing it.

Kim revealed that OpenAI is using autonomous AI agents to design the very chips meant to run future models. “The OpenAI lead chip architect told me the agents have been a dramatic multiplier and accelerator in terms of their chip design,” (play from 26:07) Kim said.

Designing hardware traditionally takes hundreds of engineers years of manual simulation, floorplanning, and physical verification. By deploying autonomous agents into that loop, OpenAI shrinks the bring-up cycle. This creates a closed recursive cycle: today's frontier models help write the hardware specifications for tomorrow's accelerators.

Meanwhile, consumption shows zero signs of plateauing. John Coogan noted that recursive self-improvement feeds directly into hardware volume: “The more advances you get, the more chips you need to buy.” (play from 25:14) Meta and OpenAI executives told Kim the exact same thing: they are taking every single chip they can acquire.

What to Do With This

Audit your compute infrastructure this week. If your startup relies on a tier-two neocloud for mission-critical training workloads, verify their balance sheet and contract termination terms before locking in an annual commitment. Run a test deployment on AWS Trainium or Google Cloud TPUs within five days to ensure your training pipeline is not married to a single vendor or hardware architecture.