AI Engineer
← All speakers

Related work and projects

Bio, Work & Ideas

Oliver Kennon

Conference affiliation: SemiAnalysis

Oliver Kennon writes about AI infrastructure finance for SemiAnalysis, an independent AI and semiconductor research firm. His work examines how computing capacity is financed and how that financing distributes risk.

In a 2026 analysis with Daniel Nishball and Terence Ong, Kennon examined how Nvidia’s commitments can help cloud operators finance GPU clusters. Under its AI Cloud Partner program, Nvidia offers to rent capacity at a below-market rate if an operator’s other customers fall away. That revenue floor can make a cluster more attractive to lenders while Nvidia earns from the initial GPU sale and potentially shares in revenue above the floor.

The authors also considered how residual value guarantees might support future financing. In their proposed structure, investors would absorb initial losses, while Nvidia’s guarantee could help protect senior lenders if equipment had to be sold for less than the debt outstanding. They weighed the capacity these arrangements could unlock against a correlated risk: a downturn could leave cloud operators unable to meet their commitments just as Nvidia’s own business weakens.

Read the topics behind these talks

1 conference talk

▶ Watch ↗

17:42

Same GPU, 50% More Output: Why the Operator Matters — Oliver Kennon, SemiAnalysis

Read the full talk →

Key ideas

Scroll to read ↓

Oliver Kennon of SemiAnalysis treats compute as a commodity whose useful output depends on serving software, production operations, security, and cost discipline—not merely the chip or its hourly rental price.

  • GPU-hours measure an input. The better economic metric is full cost per successfully completed task at the required quality; comparable price per token is only a practical proxy.
    1:39 ↗
  • In the presented B200 comparison, the same model, token mix, and interactivity target yield about 200 versus 300 output tokens per second because the serving stacks and operating choices differ.
    3:23 ↗
  • Useful intelligence requires quality, latency, availability, and security. Fast GPUs do not rescue a service that exposes another customer’s logs or cannot sustain production operation.
    4:53 ↗
  • Rental price, production cost, and selling price are separate numbers. Operations create an advantage only when their added cost is justified by more acceptable output and the process remains effective as workloads change.
    7:57 ↗
  • Compute grades and operator ratings can support credible rental benchmarks and futures, but a rental-price hedge cannot repair deployment failures or cover every component of inference production cost.
    12:41 ↗
  • Cost position is relative. Improving your own operation is insufficient when competitors lower the cost of comparable useful work faster.
    9:50 ↗

References