OpenAI reported that o1 solved 74% of 2024 AIME problems with one sample, 83% using consensus among 64 samples, and 93% with reranking of 1,000 samples. Source
1
The results establish a reason to spend more computation on difficult answers. Sample counts do not establish corresponding dollar-cost multiples.
A fixed-benchmark inference price reaches $0.07 per million tokens
Stanford’s comparison records a minimum query price of $0.07 per million tokens at its specified GPT-3.5-equivalent MMLU threshold by October 2024, down from $20 in November 2022. Source
AWS cuts the on-demand price of an existing H100 instance
A September 2025 CoreWeave disclosure describes an arrangement obligating NVIDIA to purchase residual unsold capacity, subject to delivery and availability, through April 2032. Source
Google reports a surge in aggregate token activity
At Google I/O 2026, Sundar Pichai reported more than 3.2 quadrillion monthly tokens across Google’s surfaces, versus roughly 480 trillion at the preceding year’s I/O. Source
CoreWeave reports the financial burden of expansion
CoreWeave reported second-quarter revenue of $2.575 billion and adjusted EBITDA of $1.510 billion, alongside a $49 million GAAP operating loss and $640 million of net interest expense. Source
'/%3e%3cpath%20id='Vector'%20d='M20.7611%2018.7207C20.7611%2018.7207%2017.3365%2036.1152%2028.2392%2043.9638L10.7517%2043.9777C10.621%2031.3823%2015.0227%2027.7579%2020.7611%2018.7207Z'%20fill='%23242121'/%3e%3c/g%3e%3cdefs%3e%3clinearGradient%20id='roadstar-logo-gradient'%20x1='22'%20y1='0'%20x2='22'%20y2='44'%20gradientUnits='userSpaceOnUse'%3e%3cstop%20stop-color='%23242121'/%3e%3cstop%20offset='1'%20stop-color='%23242121'%20stop-opacity='0.4'/%3e%3c/linearGradient%3e%3c/defs%3e%3c/svg%3e){kind=small}
⌘ K
By z· 2,698 words
Why compute might get 10x+ more expensive in coming years?
OpenAI shows how additional inference improves answers
[1]OpenAI reported that o1 solved 74% of 2024 AIME problems with one sample, 83% using consensus among 64 samples, and 93% with reranking of 1,000 samples. Source
The results establish a reason to spend more computation on difficult answers. Sample counts do not establish corresponding dollar-cost multiples.
A fixed-benchmark inference price reaches $0.07 per million tokens
[3]Stanford’s comparison records a minimum query price of $0.07 per million tokens at its specified GPT-3.5-equivalent MMLU threshold by October 2024, down from $20 in November 2022. Source
AWS cuts the on-demand price of an existing H100 instance
[2]AWS announced a 44% on-demand price reduction for its P5 instance using H100 GPUs. Source
NVIDIA agrees to backstop unsold CoreWeave capacity
[28]A September 2025 CoreWeave disclosure describes an arrangement obligating NVIDIA to purchase residual unsold capacity, subject to delivery and availability, through April 2032. Source
Google reports a surge in aggregate token activity
[5]At Google I/O 2026, Sundar Pichai reported more than 3.2 quadrillion monthly tokens across Google’s surfaces, versus roughly 480 trillion at the preceding year’s I/O. Source
CoreWeave reports the financial burden of expansion
[11]CoreWeave reported second-quarter revenue of $2.575 billion and adjusted EBITDA of $1.510 billion, alongside a $49 million GAAP operating loss and $640 million of net interest expense. Source
Sources
Comments