We scan new podcasts and send you the top 5 insights daily.
Evan Conrad says early H100s and GB200s broke or had firmware bugs, so sophisticated buyers may pay more for units deployed after a certain date to avoid first shipments. That cohort risk becomes another tradable property in SF Compute's order book, alongside chip generation and storage.
Trading compute futures requires more than tracking hardware supply. Software advancements, like model compression and optimization, can dramatically alter the utility and demand for older chips. A trader must understand how the software layer can make legacy hardware more capable over time, fundamentally changing supply-demand dynamics.
Unlike typical computer hardware that depreciates rapidly, H100 GPUs are trading above their launch price in secondary markets. This market anomaly, driven by the extreme and sustained compute shortage for AI, completely inverts traditional financial models for hardware assets.
Despite the rollout of Blackwell and the announcement of Vera Rubin, cloud provider Nebius reports that pricing for older Hopper GPUs is not dropping, and in some cases is rising. This shows a persistent market for "good enough" compute for specific workloads.
Don Wilson says the trillion-dollar question is what a B200 will be worth four or five years out. H100 on-demand prices fell for months, then rose after Anthropic's new model boosted inference demand, while token prices moved differently. That volatility is why he argues forward contracts are badly needed.
Contrary to typical hardware depreciation, GPUs like NVIDIA's H100 are becoming more valuable over time. This is because newer, more efficient AI models can generate significantly more output and value on the same hardware, tying the GPU's worth to its utility rather than its age.
Evan Conrad says two VMs from one operator can differ by 40% in kernel-launch performance, and storage, WAN transit, preemption and chip generation also change value. Rather than wait for standardized contracts, SF Compute lets buyers and sellers attach multiple properties to orders, accepting wider spreads and arbitrage for flexible buyers.
The mean time between failure for GPUs depends heavily on their use. Chips used for intensive model training have much higher failure rates than those used for inference. Investors often conflate the two, underestimating the high churn and replacement costs for training hardware.
Counter to narratives about rapid depreciation, the market for used high-end GPUs is robust. Data from late 2023 showed a second-year H100 reselling for 85 cents on the dollar, and a third-year for 84 cents. This high residual value makes refurbished chips a viable and capital-efficient option for compute providers.
A futures market for GPU compute is not viable yet because the product isn't fungible. The performance of an identical H100 chip varies significantly between cloud providers based on their proprietary software stack and operational excellence, measured by metrics like "goodput" and "MFUs."
The concept of GPUs as a fungible commodity is complicated by significant performance differences between identical chips. Research on A100s shows up to 38% variance due to chip-level and provider differences. This necessitates verification services to ensure buyers get the performance they pay for, challenging the idea of perfect interchangeability.